Review guidance

• This manual is aimed at aspiring and professional conservationists who want to understand and use AI effectively and ethically.
• We will put in citations and references once all content is finished - but please feel free to suggest important reference materials.
• Otherwise you can put in comments or make suggested edits as you like.
• Would be great if you leave a comment saying your name and association (e.g. Adam Barlow, WildTeam), so that we can acknowledge your review contribution.

AI for Wildlife Conservation

Acknowledgements

Graham Wallington, Xeroth AI

Jessica Hodgson, Independent

Julia Li, TUM

Karina Torres, Southern Illinois University

Md Anwar Hossain, Independent

Naomi Matthews, Chester Zoo

Yanna Vogiazou, Independent consultant

Introduction

Background

Artificial Intelligence (AI) (Box 1) is a tool that speeds up a huge range of tasks, so you can get more done in less time. This has huge implications for your conservation work, because AI can draft emails, ^[a]summarise reports,^[b] and analyse data in a fraction of the time it would take you on your own. Combined with other technologies, AI can now^[c] also help identify species from camera trap images, track habitat loss from satellite imagery, and detect poaching activity.

The Conservation AI^[d] platform now processes millions of images^[e] each week across reserves in Uganda, Kenya, and South Africa, using real-time^[f] machine learning models that automatically classify^[g] hundreds of wildlife species. Beyond wildlife tracking, these smart camera networks detect weapons, unauthorized vehicles, and poachers in real time^[h], to trigger immediate ranger deployments and generate evidence for courtroom convictions^[i]. Meanwhile, in the Peruvian Amazon, platforms like Forest Foresight^[j] use AI to model satellite data, illegal road building, and environmental indicators to pinpoint high-risk rainforest zones up to six months before clearing^[k] begins, enabling authorities to intervene before a single tree falls.

As well as specialist conservation uses, AI is becoming fully integrated into conservationists use of the internet, applications, and software in general. For example, Microsoft^[l][m] has built its Copilot assistant into Office applications such as Word, Excel, and Outlook, and into the Edge browser, putting AI assistance inside the tools conservationists already use for everyday tasks. ^[n]

This means conservation teams are likely to change in size and in the work they do, which will affect every conservationist's career. People who build their AI skills will be able to do more at the same time. ^[o]For example, a monitoring officer who can use a species identification model can process a season of camera trap images in days instead of months. These skills also open up new roles. An organisation may need a Head of AI Application to set its policy on responsible AI use and train staff to apply it, or a data officer who can connect AI agents to monitoring databases. Building AI skills now will put you in a stronger position for these opportunities. ^[p][q]

However, this change also means that fewer people may be needed to do the kind of work that a conservationist used to do^[r], like collecting data^[s][t], creating social media posts, and drafting funding proposals. For example, a large NGO may be able to raise the same amount of funds with one staff member using AI, that they were previously able to do with a team of four. The purpose of this best practice, therefore, is to help conservationists understand what AI is, and to provide guidance on how to use it effectively and responsibly in your conservation work.

Application^[u]

You can use this AI for Wildlife Conservation best practice whenever you are working out whether^[v][w] or how to use AI in your conservation work. You can use it as:

• A personal guide to check your own use of AI against, before you start a task.
• The basis for an organisational policy on responsible AI use.
• A training resource for building AI capability across your team or network.
• A reference for assessing whether a new AI tool is fit for a specific task.

How closely you apply it depends on what you are producing and which AI tool you have chosen. A quick social media post calls for a lighter check than a funding proposal a donor will scrutinise, but the same principles and process apply to both.

Overview

This best practice is made up of the following componants:

• Understanding AI: What AI is, how AI models are built, how AI systems and AI agents work, and how AI is likely to change, so you can judge what it is capable of and where it can go wrong.
• Principles: Six principles that should run through every use of AI, covering your own expertise, moderation, bias, accountability, the value of what you produce, and openness about AI's part in it.
• Using AI: A process for producing one piece of work with AI, made up of designing the output, selecting a tool, setting its permissions, instructing it, verifying what it produces, and finalising the piece.

While Using AI is set out as a sequence, it is not always followed in a straight line. New information at any step can send you back to an earlier one, such as a source that contradicts a draft or a tool that turns out to be unsuited to the task.

Supporting resources

TBC

Understanding AI

Overview

For many people, AI is close to a black box. You type a prompt, or a camera trap photo gets identified, and a result appears, with little sense of what happened in between^[x]. That gap makes it easy to hold onto a picture of AI that is not quite right, such as treating it as something that thinks or reasons the way a person does. AI is also being built into more of the everyday tools you already use, from email to field sensors, often without you noticing. A clear picture of what AI is and where it is heading gives you the foundation for understanding the principles and process for using it.

This section builds that picture, covering:

• What AI is, describing the components of an AI system and the two main types of AI model, predictive and generative.
• How AI models are built, setting out the steps developers follow to create a model, from defining its purpose to releasing it.
• How AI systems work, explaining how simple AI systems and AI agents turn an input into an output or action, how permissions limit what AI agents can do, and where AI systems run.
• How AI will change, describing how AI is doing more without a person and being built into more digital tools and physical devices.

What AI is

AI is an area of computer science focused on developing AI systems^[y]. An AI system is a computer or machine system that uses AI to perform tasks that normally require human intelligence, uch as analysing information, making predictions, writing a report or creating a picture. In general, an AI system can be made up of the following components^[z]:

• Data:The information built into and used by the system. It is a core component because it provides the knowledge to build and run the AI. This includes the massive^[aa] datasets used to train the system’s memory, as well as new information fed in to generate answers. For example, thousands of camera-trap^[ab] photographs^[ac] from a wildlife reserve could be used to train the model or provide images for it to analyse.
• Model: the trained computational model that provides the AI capability^[ad] (Box X). For example, a model trained to recognise animals in camera-trap photographs could identify whether an image contains a leopard, elephant, or other species.
• Application: the software that uses the AI model to perform a particular task and provides a way for people to interact with the system. For example, a camera-trap monitoring application could automatically send photographs to the AI model, record the species identified and display the results for a conservationist to review.
• Tools: an AI system may connect to other software or equipment to access information or carry out tasks. For example, the camera-trap system could connect to a database to store observations, a mapping system to show where animals were detected, or email software to alert a conservationist when a leopard is detected.

Figure X. AI system components.

Box X. Types of AI model.

Broadly, there are two common types of AI model:

Predictive models analyse input data and use patterns learned from existing data to produce an output such as a classification, prediction or detection. For example, a model could identify animals in photographs, recognise bird species from their calls, classify different types of habitat or predict where a species is likely to occur.

Generative models learn patterns in existing data and use them to generate new content. They can produce text, images, audio, video or code. Large language models (LLMs) are a type of generative model designed to work with language. Tools such as ChatGPT use LLMs to generate and analyse text.

Not every AI system contains all of these components. One AI system might consist of an AI model within an application, while a more complex system might use several AI models, databases and software tools to carry out a series of tasks.

How AI models are built

Every AI model is built through the same broad sequence of steps^[ae], whether it identifies bird calls or writes text. The steps differ mainly in scale. A predictive model for one conservation task can be built by a small team of developers, while a large generative model takes years of work, thousands of specialised computer chips, and a budget that few developers can afford.

You do not need to build a model to use AI well, but knowing how models are built explains most of their limits. It shows why mistakes in the training data become mistakes in the model, why a bird call identification model can fail on recordings from a different region, why a chatbot can repeat the biases in its training data, and why a model can be out of date. Understanding these limits will help you choose the right AI tool for a task and know what to check in its outputs.

The steps for building an AI model are shown in Figure X and outlined in more detail below.

Figure X. Steps for building an AI model.

Define purpose

The developers first decide what the model needs to do and how they will judge whether it does it well. For a predictive model this task is narrow, such as identifying 30 mammal species in camera trap photographs from one region. For a generative model it is broad, such as producing useful text in response to almost any prompt. This decision shapes every step that follows.

Design model^[af]

Developers write the code that sets out the model's structure and how it will learn. The task decides that structure. A model that identifies species in camera trap photographs needs a structure suited to images, while a model that writes text needs a structure suited to language. The task also sets the model's rough size, so a model that identifies 30 species is far smaller than a model that responds to almost any prompt.

Most modern models are neural networks, which are layers of connected calculations loosely inspired by the brain. The code sets out what those calculations are and the order they run in, and it stays the same once written. The calculations also use a set of numbers called parameters, which are stored in a separate file from the code and can be changed. A model can have millions or billions of parameters, and their number is the usual measure of a model's size. At this stage the parameters are set at random, so the calculations produce nonsense and the model can do nothing useful.

A mixing desk in a recording studio works in a similar way. The code is the wiring of the desk, which fixes how the sound flows through it. The parameters are the positions of the knobs. Training, covered in Step 4, turns those knobs until the output is right.

Prepare training data

Developers then collect the training data the model will learn from. A predictive model needs training data for the one task it will do, such as camera trap photographs, satellite images of cleared forest, or recordings of whale calls. A generative model needs far more training data, gathered from across the internet along with books, articles, and images that developers license. Whether the people who created that training data agreed to its use, or were paid for it, is an open legal and ethical question, covered under ethical considerations in Select tool.^[ag]

For a predictive model, people give each photograph or recording a label showing the correct answer^[ah], such as which species appears in a camera trap photograph. This labelling is slow and often done by experts or citizen science volunteers, and any mistakes in the labels become mistakes in the model.

The training data for a generative model is too large to check by hand, so developers use automated filters to remove duplicates, spam, and some harmful content. Nobody checks it for accuracy^[ai]. It keeps reliable content alongside content that is outdated, biased, or wrong, and this is the source of the bias covered in the See past the bias principle.

In both cases, developers set aside part of the training data as test data. The model never learns from the test data, and developers use it in Step 5.

Train model

Training is where the model learns. The model is shown one item from its training data, makes a prediction, and compares that prediction with the correct answer. For a predictive model, the correct answer is the label. For a generative model, it is the next word in a real piece of text, which is why generative models can learn from text nobody has labelled.

After each comparison, the parameters are adjusted slightly so the next prediction comes a little closer, while the code stays the same. Repeated billions of times, these small adjustments add up to a model that has learned the patterns in its training data. Nobody writes a rule describing a leopard's rosettes, because the model learns that pattern from labelled photographs.

Developers often follow this with a second, smaller round of training called fine tuning. A generative model trained only to predict the next word can continue a piece of text, but it does not yet follow prompts well. Developers therefore fine tune it on outputs written by people to show what a good output looks like, and on people's ratings of which of two outputs is better. Fine tuning teaches the model to follow prompts and to decline some harmful prompts.

Training a large generative model runs for weeks or months on thousands of specialised computer chips in data centres, using large amounts of electricity and water. This is part of the cost covered in the Use with moderation principle.

Test model

Developers test the model on the test data they set aside in Step 3, to measure how often its predictions are correct. A model can perform well on its training data and still fail on new data, so this test gives a truer picture of how it will perform once released^[aj].

Testing also looks for weak spots. A model trained on camera trap photographs from East African savanna may perform well there and fail on camera trap photographs from a Southeast Asian rainforest, where the species and lighting differ. ^[ak]Developers also put generative models through safety testing, in which people try to make them produce harmful or false content.

Where the model falls short, developers improve it by returning to earlier steps, most often to gather more training data or change how the model is trained, and then test it again. Safety testing also checks how a model behaves once it is built into an AI agent, such as whether it takes unexpected shortcuts when it meets an obstacle (Box X).

Release model

Once the model is released, its parameters are fixed.^[al] The released model is the code together with the file of trained parameters, and it is built into applications such as a camera trap monitoring system or a chatbot. The model stops learning at this point. Some AI tools save notes about you between conversations or search the internet for recent information, but neither changes the model itself. Only a new version, built by repeating these steps with newer training data, brings the model's knowledge up to date, so a model can be confidently out of date.

How AI systems work

Every AI system takes an input, passes it to one or more trained models, and turns what those models produce into an output or an action. Both types of AI system can use tools to do this. A tool can be digital, such as a database or email software, or physical, such as a drone, a robot, or a vehicle.

How an AI system works depends on two things. The first is who decides the steps in between. In a simple AI system, a person sets those steps in advance, and in an AI agent, the AI agent decides them for itself as it goes. The second is where the AI system runs, either in a data centre or on the device that captures the input^[am], which affects how its output reaches you.

Simple AI systems

A simple AI system carries out a fixed sequence of steps that a person designed for one task. The model supplies the intelligence at one or more of those steps, and the application decides what happens before and after. However sophisticated the model is, the simple AI system uses the same general sequenceof steps (Figure X).

• Capture input: A sensor, a person, or another piece of software provides the input, such as a photograph, a sound recording, or a typed prompt, and the application receives it.
• Run model: The application passes the input to the model, which compares it against the patterns it learned in training and returns an output, such as a prediction with a score for how confident it is.
• Apply rules: The application checks the model's output against rules a person set in advance, such as accepting it only above a set confidence score ^[an]and sending anything less certain to a person to check.
• Produce output: The application acts on the accepted result, such as saving a record, updating a map, or sending an alert.

Figure X. Sequence of steps for a simple AI system.

For example, an acoustic monitoring system in a rainforest reserve might work as follows:

• Capture input: A microphone mounted in the canopy records a loud sound and sends the recording to the application.
• Run model: The application passes the recording to a predictive model trained on thousands of labelled forest sounds, which predicts that the sound is a gunshot and is 95 percent confident.
• Apply rules: The rules state that any gunshot prediction above 80 percent confidence is accepted, and any below it is sent to a ranger to listen to^[ao]. This prediction is accepted.
• Produce output: The application records the time and location of the gunshot, marks it on the reserve map, and sends an alert to the ranger team on patrol nearby.

A basic chatbot is also a simple AI system, built around a generative model. The input is your prompt and any files you upload. The model, usually a large language^[ap] model, then predicts the most likely next word, or part of a word, one at a time. Given "The elephant walked into the", it rates "clearing," "water," and "forest" as likely next words and "teapot" as close to impossible. It adds its chosen word to what it has written so far and predicts the next one, building up sentences and paragraphs. The output is the text that appears on your screen. AI tools that generate images and video follow the same sequence, although their models start from random visual noise and adjust the whole image over many steps until it matches the prompt.^[aq]

The output of a simple AI system can also be a physical action. A habitat monitoring drone can use a predictive model to recognise a riverbed in its camera feed, and its application turns that prediction into steering commands, so the drone follows the river and collects data on tree cover as it goes. Motors, rotors, and other engineering carry out the action. Making a device act safely in a real forest or river is a harder problem than producing text, which is why AI systems that act physically lag behind the other AI tools covered in this best practice.

AI agents

An AI agent is given a goal and decides for itself which steps will reach it. It works through the following steps:

• Receive goal: A person gives the AI agent a goal, usually as a written instruction.
• Select action: The AI agent's model, usually a large language model, works out which action will move it closer to the goal, based on the goal and everything it has found so far.
• Check permissions: The AI system checks whether the AI agent's permissions allow that action, such as reading a database but not deleting from it. If the action needs a person's approval, the AI agent waits for it.
• Carry out action: A tool carries out the action and returns the output to the model. Tools can be digital, such as databases, internet search, email, and other models, including predictive models, or physical, such as drones and vehicles.
• Assess output: The model judges whether the goal is met . If not met, the AI agent returns to Decide the next action and repeats the steps from there until it is.
• Deliver result: Once the goal is met, the AI agent produces the final output, such as a report, a saved file, or a message.

Figure X. Sequence of steps for AI agents.

For example, an AI agent assessing an elephant population might work as follows:^[ar]

• Receive goal: A reserve manager asks the AI agent to "assess whether the elephant population in this reserve has changed significantly over the last five years, identify the likely causes of any change, and produce a report".
• Select action: The model decides it first needs the survey data.
• Check permissions: The AI agent has permission to read the reserve's database, so it can go ahead.
• Carry out action: The database tool returns five years of survey data.
• Assess output: The model judges that the goal is not yet met, so the AI agent returns to Select action and repeats the steps several times. It uses a predictive model to estimate the population trend, which shows a decline. It then decides to check whether habitat loss could explain the decline, so it retrieves satellite images of the reserve and uses a second predictive model to measure habitat change. It searches the internet for recent research on elephant populations in the region, plots the population trend and habitat change with a charting tool, drafts the report with its large language model, and checks the draft against the survey data, revising any claim the data does not support. Once the report is complete and checked, the model judges the goal is met.
• Deliver result: The AI agent saves the report. Because sending files outside the reserve's system needs approval, it waits for the reserve manager to approve before emailing the report to the conservation team.

Nobody told the AI agent to check habitat loss or search for recent research. It chose those actions because of the decline it found in the survey data, so a second run toward the same goal could take a different route. ^[as]

The same difference applies when an AI system acts physically. A habitat monitoring drone running a simple AI system follows a river because a person set that sequence in advance. Running an AI agent, the same drone could decide for itself to leave the river, take a closer look at a possible clearing, and then return to its route. Its permissions would decide at Check permissions whether it can do this without a person approving it.

The wider an AI agent's permissions, the more it matters to decide in advance which actions it can take without a person approving them. Permissions sit outside the model, in the settings of the AI system and the tools it connects to. The developers of the AI system decide which permissions are available and what the defaults are. The person or organisation using it then chooses from those, for example by giving the AI agent read only access to a database or requiring it to ask for approval before it deletes anything. An instruction typed into the conversation, such as "do not change the database", is different, because the model reads it as text and nothing stops it from acting otherwise. The permissions an AI agent has therefore set the limits of the harm it can cause, although that harm can be difficult to predict in advance (Box X).

Box 3. AI unintended harm.^[at][au][av]

An AI agent does whatever its goal, tools, and permissions allow, and when it meets an obstacle it can take a route to its goal that nobody intended. In one test, AI agents told to beat strong chess software edited the file that stored the board so their opponent's pieces disappeared. In another case, an AI agent asked to move a user's files into a new folder failed to create the folder, did not check, and moved the files anyway.

AI agents can also be manipulated by instructions hidden in an email, document, or web page they read, because they cannot reliably tell those instructions apart from yours. For example, an AI agent with access to your email and monitoring database could read a hidden instruction in an email and send the GPS locations of collared rhinos to someone planning a poaching attempt.

It is also important to be aware that the same AI capabilities that help conservation can also be used deliberately to cause harm. AI can generate realistic fake images, video, and audio, such as a photograph of a ranger mistreating a suspect or a recording of a director's voice asking staff to transfer funds. These can be used to discredit a conservation organisation, spread false stories about a community, or commit fraud. AI can also be used to monitor people, for example by running facial recognition on camera trap images to identify community members who enter a protected area, which can breach their rights and damage their trust in conservation.

Where AI systems run

Most AI systems built around a generative model, such as chatbots, run on powerful computers in large buildings called data centres^[aw]. The input, such as a text instruction, travels there over the internet, the AI system works on it, and the output travels back to your screen, usually within seconds. This is why most chatbots need an internet connection, and why anything that is typed into them leaves the device.

Ai systems can also run on the device itself. This is called edge AI. A camera trap with edge AI can identify a species on the spot and send only the prediction, or an alert, over a satellite or mobile link. ^[ax]This suits remote field sites, since the device keeps working without a reliable internet connection and uses less power sending data. Sensitive photographs, such as those showing people, can also stay on the device. The trade off is that an AI system running on a device in the field has to use a smaller model, which is usually less capable than a model in a data centre.

How AI will change

What is changing fast is that AI is:

• Doing more without a person: Programs increasingly plan out a series of steps, carry them out using other software, and keep going without waiting for approval at each step. For conservation work, that could mean a programme that manages a whole grant application from first draft to submission, or one that runs a habitat monitoring programme from end to end, starting with collecting the data and ending with sending a finished report to stakeholders.
• Being built into more digital tools: AI is likely to stop being something you open in a separate app or window. Instead, it will show up built into what you already use, email, spreadsheets, mapping tools, and field data collection apps. The same shift is likely across a conservation organisation as a whole, with AI built into donor databases, grant management systems, and monitoring dashboards, instead of sitting in one programme a few people know how to use.
• Being built into more physical devices: AI is increasingly being built into physical devices such as drones, camera traps, and field sensors, alongside the associated software on a laptop or phone. For example, a drone can already plan its own flight path, spot signs of illegal logging as it flies, and adjust course to get a closer look, without a person piloting it step by step. Camera traps and acoustic sensors are moving the same way, processing and acting on what they capture in the field instead of only recording it for someone to review later.

While these changes may help us do more faster, the increased sophistication, pervasiveness, and autonomy of AI may also increase the risks of unintended harm..

Principles

Overview

Used without any guardrails, AI can waste resources on a task that never needed it^[ay], or reproduce the bias in its training data without anyone catching it. It can produce confident answers that nobody involved knows enough to check, and work that carries no one's name and answers to no one.^[az] It can also add to a pile of generic content that gives little back to whoever reads it, or leave people unaware that AI played a part in what they receive^[ba][bb]. The purpose of the principles, therefore, is to help you use AI effectively and ethically across the whole of your conservation work, regardless of what you are producing or which AI tool you have chosen.

The six principles are:

• Know enough to judge it, to keep your use of AI within what you can check yourself while you keep building your own expertise.
• Use with moderation, to use AI only where it suits the task, and in proportion to what the task needs.
• See past the bias, to catch the different kinds of bias AI can reproduce from its training data before AI generated content goes out.
• Keep people in the loop, to keep a named person responsible for anything AI helps create or carry out.
• Skim off the slop, to keep anything you produce with AI worth the time it asks of the person reading it.
• Be open about AI, to make sure people know when AI played a substantial part in something they receive.

Principles should not be confused with the process steps that follow them. While the process steps describe the specific actions you take to produce one piece of work with AI, from designing what you are making through to finalising it, principles describe the overarching approach and understanding that should run through all of it, whichever action you are taking and whatever you are producing.

Know enough to judge it

Purpose

Judging whether an AI output is any good takes the same knowledge and skill that the AI is standing in for. Generative AI can produce a wrong answer that sounds as confident as a right one, so relying on it for work you could not do yourself leaves you unable to tell the two apart. Relying on it too early can also stop you building that knowledge and skill in the first place, which matters most for people early in their conservation careers. The purpose of the Know enough to judge it principle is to keep your use of AI within what you can check yourself, and to keep building your own expertise while you use it.

Application

The Know enough to judge it principle can be applied by:

• Using AI only for work you know enough to check. ^[bc]Where a task goes beyond your own expertise, ask someone who has that expertise to review the output before you use it. For example, a project officer asks AI to suggest a sampling design for a bat survey, then has the organisation's ecologist check it before any fieldwork is planned, because the officer has no training in survey design.
• Forming your own view before asking AI. Work out your own answer or position first, then use AI to test or extend it, so its output does not simply replace your thinking. For example, a site manager writes her own list of the main threats to a wetland before asking AI to identify threats, then compares the two lists and investigates the differences.
• Doing tasks yourself while you are still learning them. Where a skill is central to your role, do the work without AI until you can do it well, and use AI to check or speed it up afterwards. For example, a new fundraising assistant writes his first grant applications himself with feedback from his manager, and only starts drafting with AI once he can spot a weak application.

Use with moderation

Purpose

Every use of AI runs on computer hardware^[bd], which draws electricity and water and eventually becomes waste. Heavier tasks, such as generating an image or a video, use far more of this than a short written answer. The purpose of the Use with moderation principle is to keep AI use in proportion to what a task  needs, so it does not create environmental or resource cost the task does not justify.

Application

The Use with moderation principle can be applied through:

• Checking whether AI is the right tool for the task at all. Some tasks are quicker, more reliable, or more effective without AI. For example, a data officer calculating monthly totals from patrol records uses a spreadsheet formula, which gives the same correct answer every time, and a volunteer coordinator writes thank you notes to long serving volunteers herself, because the personal effort is the point of the note.
• Match the weight of the AI tool to what the task needs. For example, a communications officer needing some kind of image for an internal briefing keeps to a simple image tool at low resolution, rather than generating a three minute video.
• Check whether something real already exists before generating something new. For example, a fundraiser spends 20 minutes finding an existing camera trap photo of an elephant corridor rather than generating an AI image in seconds.
• Settle on a result early instead of regenerating it over and over for small changes. E.g. a designer refining a fundraising poster picks a strong AI generated draft, then makes the remaining small changes using graphic design software.

See past the bias

Purpose

An AI programme learns its patterns from a huge collection of existing material, and that material carries the biases already present in the world it came from. The programme has no sense that any of this is biased, so it reproduces the most common pattern in its training data the same way it reproduces anything else it generates.This means if the information base shows leaders are commonly white, middle-aged men, then that will be reflected in the outputs it creates. The purpose of the See past the bias principle, therefore, is to enable conservationists to be aware of, identify, and account for AI bias before generated content is used.

Application

You can apply the See past the bias principle through:

• Checking output for representation bias.  For example, a comms officer generating illustrations for a training brochure on rangers notices every default output shows a white man, and asks explicitly for a mix of genders and ethnicities that reflects the real team.
• Checking output for expertise bias. For example, a programme lead reviewing an AI drafted report on traditional fire management notices it cites international academic sources as evidence throughout, while summarising local land managers' input as anecdote, and rewrites it to credit their knowledge as expert testimony in its own right.
• Checking output for worldview bias.  For example, a training officer drafting community engagement materials for a conservation programme notices the AI's suggested framing treats individual land ownership and formal legal titling as the obvious solution, and rewrites it to reflect the community's own communal land tenure system.
• Check output for political bias. For example, a policy officer drafting a briefing for a cross party parliamentary group notices the AI's language frames stronger environmental regulation as the only reasonable position, and rewrites it with a range of options, each assessed using a cost-benefit analysis.

Keep people in the loop

Purpose

As AI takes on more of the work itself, planning a series of steps, carrying them out, and acting with less oversight along the way, it becomes easier for no one in particular to be answerable for what it produces or does. The purpose of the Keep people in the loop principle is to keep a named person responsible for anything an AI programme helps create or carry out, since the programme itself cannot be held accountable for it, and someone needs to be able to explain, defend, or correct an output made with its help.

Application

The Keep people in the loop principle can be applied^[be] by:

• A person taking ownership of anything they produce using AI. For example,  a Director signs off personally on an annual impact report drafted with AI, instead of sending it out under the organisation's name with no one accountable for its content.
• A person deciding in advance which decisions an AI programme can make. For example, a ranger team lead sets a habitat monitoring drone to alert them and wait for further instructions whenever it detects a person in a restricted zone.

Skim off the slop

Purpose

Readers are getting better at recognising AI generated text and images as more of it appears. A page of AI text that nobody has shaped, checked, or written adds to a wider slop of interchangeable content that is harder to trust and harder to enjoy, wherever it turns up. The purpose of the Don't add to the slop principle is to make sure anything you produce with AI is still worth the time it asks of the person who will receive and digest it.

Application

The Don't add to the slop principle can be applied through:

• Learning to recognise typical AI output, including the generic phrasing, terms, structure, and vague, upbeat tone, so you can catch it in your own drafts before a reader does. For example, an education officer recognises the AI slop in the following output “In today's rapidly evolving world, climate change isn't just a challenge — it's an opportunity to genuinely come together and make a difference. At the end of the day, it's not just about saving the environment — it's about safeguarding our shared destiny.” is AI slop (Box X, Figure X).  
• Prompt against generic AI style before you start, by instructing the tool explicitly what to avoid and giving it an example of the tone, structure, and voice you want it to match. For example, a fundraiser drafting a newsletter appeal tells the AI to avoid clichés and rule of three lists, then pastes in last year's best performing appeal to reference for tone, terminology, and structure.
• Adding in something that the AI could not generate, such as a specific detail, example, or judgement. For example, a programme officer swaps the AI's generic closing paragraph in a funding update for a specific detail from last week's site visit.

Box X. Examples of AI text slop.

Funding appeal: "Protecting this rainforest isn't just about saving trees, it's about safeguarding a future for the wildlife that genuinely depends on it."

Social media caption: "Every donation truly makes a difference for elephants like these! #WildlifeConservation #ProtectOurPlanet #EverySpeciesMatters"

Report conclusion: "The pilot programme delivered strong results — better data collection and genuinely fewer conflict incidents. Moving forward, these lessons should shape how the next phase is designed."

Figure X. Examples of AI image slop.

Be open about AI

Purpose

People judge information differently depending on who, or what, produced it. A funder reading a proposal, a community member reading a leaflet, or a supporter looking at a campaign photograph may reasonably assume a person wrote or captured it, and finding out later that AI played a large part can damage their trust in you and your organisation. Some funders, journals, and partners also require AI use to be declared. The purpose of the Be open about AI principle is to make sure people know when AI played a substantial part in something they receive, so they can judge it accordingly. Routine use, such as checking spelling or tidying the wording of an internal email, does not need to be flagged.

Application^[bf]

The Be open about AI principle can be applied by:

• Deciding whether the audience would want to know. Consider how much of the piece AI produced, whether the audience would assume a person made it, and whether they have their own rules on AI use. For example, a programme manager checks a funder's guidance before submitting a proposal drafted with AI, finds that it requires AI use to be declared, and adds a statement to the application.
• Saying plainly what AI did and what a person checked. Name the AI tool, what you used it for, and how the output was reviewed. For example, a research officer adds a note to a species status report stating that AI was used to summarise the literature review, and that all cited sources were checked by the author.
• Labelling AI generated images and video that could be mistaken for real. For example, a communications officer adds the caption "AI generated illustration" to an image of a pangolin in a wildlife market used in a campaign post, so supporters do not take it for evidence of a real incident.
• Telling people when they are communicating with AI. For example, a community liaison officer setting up an automated WhatsApp service for reporting crop raiding by elephants makes sure the first message states that replies are generated by AI and explains how to reach a person.

Using AI

Overview

AI tools are getting easier to use by the day, but a quick result is not automatically one that fits your audience, gets the tone right, or stands up once someone checks the information in it. This section sets out how to create a quality output through the following steps:

• Design output, deciding what you are making and what a good version of it looks like before you start.
• Select tool, picking the tool that suits the task, judged on more than convenience.
• Instruct tool, giving the AI context and clear instructions, then building on what it produces.
• Verify, checking what it produced against your sources and other evidence.
• Finalise, correcting what verification turned up and editing until it reads the way you intend, and disclosing the AI assistance where appropriate.

These steps are shown in Figure X and detailed below.  

Figure X. Steps for using AI.

Design output^[bg]

To design the output you need to first identify the audience and then decide on the format of the output.^[bh]

• Identify the audience: Name who will read or use the finished output, since that shapes everything that follows, including the level of detail, the tone, the length, and what information is presented. For example, a report written for a technical donor with an ecological background may require scientific terminology and charts of supporting data.
• Decide the format: Settle on what kind of output is required, such as a report, an email, a slide deck, a social media post, or something else. The format affects the length, structure, and tone an output needs, and it determines which AI tool is suited to the job.

Select tool

New AI tools appear constantly, and those that are available today will not be the same ones available in a year. You can use the following set of criteria to help you select which tool to use:  

• Capability: Check the tool  produces the kind of output you need, and how well it performs that task. For example, a communications manager comparing two chatbots for translation work tests both using the same awkward, idiomatic sentence to see what it produces as an output.
• Ease of use: Consider how much time and skill it takes to get a usable result. A powerful tool that takes hours to learn may not be worth adopting for a task your team does once a year.  For example, a field team without a design background picks an image generation tool activated by simple prompts.
• Cost: Weigh the price of the tool against how often you will use it and what the task is worth. For example, an organisation creating one video explainer a year picks a lower cost tool with a free tier.
• Ethical considerations: Look at how a tool was built, including whether it credits or compensates the people whose work trained it, and what its running cost is in energy and water. For example, a comms team choosing between two similar image generators checks how each sources its training images before picking one for a public campaign.
• Terms and conditions: Read the tool's terms and conditions to find out what happens to anything you put into it, and check they are compatible with your organisation's data policy. Look for whether the provider stores your inputs and for how long, whether it uses them to train future models, and whether you can switch that off. Check also who owns what the tool produces and whether you are allowed to use it commercially. For example, a monitoring manager rules out a free chatbot for summarising patrol reports after finding its terms allow the provider to use inputs for training, and chooses a paid version whose terms exclude this. ^[bi]

Set permissions^[bj]

Before you start using an AI tool, set what it can access and what it is allowed to do. This matters most for AI agents and for any AI tool connected to your own files, email, or databases, since the wider its permissions, the more harm it can cause when something goes wrong (Box X). Permissions are set in the settings of the AI tool and the tools it connects to. An instruction typed into the conversation, such as "do not change the database", does not limit what the AI tool can do.

You can set permissions by deciding:

• What it can access: Connect the AI tool only to the files, folders, and databases the task needs. For example, a researcher using an AI agent to analyse bird survey data connects it to the survey folder and leaves the rest of the shared drive unconnected.
• What it can do: Give the AI tool the lowest level of access the task needs, such as permission to read a database without permission to change or delete anything in it. For example, a monitoring officer connecting an AI agent to the camera trap database gives it read only access and turns off its ability to send email.
• What needs approval: Require a person to approve any action that cannot be undone or that sends information outside your organisation. For example, a grants officer lets an AI agent draft and save a funding application but sets it to wait for approval before submitting it to the funder.

Instruct tool

Instructing the tool to produce an output involves providing the tool with context, and providing the tool with instructions.

Give it context

Within the AI tool, you should upload or link to your source material with information that you want the AI output to use. For example, a fundraiser drafting a grant narrative pastes in the project's monitoring data and last year's approved proposal.

Likewise, you should upload or link to a strong example of the output type you want it to generate, ideally one from your own organisation or a similar one. For example, a grant manager drafting a donor update pastes in the organisation's best performing newsletter from the previous year.

Provide instructions

Use a prompt (Box X), to instruct the AI tool on such things as:

• Purpose of context materials
• Audience
• Length or dimensions
• Information sources
• Style
• Format
• Structure
• Process

Box X. Prompt.

A prompt is the instruction you give an AI tool to tell it what to produce, typically typed into a chat window or a piece of software built around one, e.g. "write a two paragraph update for our donor newsletter about the start of the new sea turtle hatching season". You can also ask an AI tool to help you write the prompt itself, by describing what you are trying to achieve in plain terms, and asking it to turn that into a clear, structured prompt

The amount of detail you put in your prompt is important for any type of AI output. More detail gives you a better quality first draft, closer to what you had in mind. Less detail gives the AI tool room to produce something you would not have thought of yourself, sometimes better than what you could have pictured.

Example grant proposal prompt

• Purpose: Draft the first version of a funding proposal for the Cairngorms Peatland Restoration Project.
• Audience: A programme officer at a funder with no ecology background.
• Length or dimensions: Three pages, under 1,500 words.
• Information sources: The attached monitoring data and project budget.
• Style: Match the tone of the proposal that secured our last grant from this funder; avoid technical jargon.
• Format: Written funding proposal.
• Structure: Match the structure of that same successful proposal; cover the problem, our approach, the budget, and the expected outcomes, in that order.
• Process: Do not state a total budget figure until the outcomes have been explained.

Example campaign graphic prompt

• Purpose: Encourage members of the public to report suspected badger baiting to the police or the RSPCA.
• Audience: Dog walkers and countryside visitors in rural England who may witness activity without recognising it as a crime.
• Length or dimensions: A3 poster.
• Information sources: The attached river cleanup campaign poster, for colour palette and logo placement.
• Style: A realistic badger set scene instead of a cartoon style.
• Format: Poster style graphic.
• Structure: A headline, the reporting number, and a short line naming the signs to look out for, such as dogs with fresh injuries, spades, or terrier collars near a set, positioned so it reads clearly from a noticeboard.
• Process: Do not show any violence toward the badger itself.

Example data analysis prompt

• Purpose: Identify patterns in human-wildlife conflict incidents to inform next quarter's patrol routes.
• Audience: The ranger operations manager.
• Length or dimensions: Two pages.
• Information sources: The attached spreadsheet of incidents from the past 12 months.
• Style: Plain language, avoiding statistical jargon.
• Format: A chart, with a short written summary.
• Structure: Present as a bar chart, grouped by species and location, ranking locations by frequency.
• Process: Rank locations by frequency before suggesting any changes to patrols; do not recommend a change for a location with fewer than three recorded incidents.

Example application development  prompt

• Purpose: Create a mobile web application that enables members of the public to quickly and anonymously report suspected listings or sightings of illegal wildlife products to enforcement authorities.
• Audience: Non-technical everyday citizens who may encounter illegal wildlife items for sale online, at local markets, or in shops.
• Length or dimensions: A single-page, fully responsive web application with a maximum 3-step reporting workflow.
• Information sources: The attached CITES species threat database and national wildlife crime reporting guidelines.
• Style: Clean, highly intuitive, and accessible interface (WCAG AA compliant) with reassuring security indicators to build user trust around anonymity.
• Format: Interactive web application.
• Structure: A prominent emergency notice/safety reminder at the top, followed by a 3-step reporting form (1. Product & Species Details, 2. Location or Web URL, 3. Photos & Narrative), ending with a clear submission confirmation screen containing a unique tracking reference code.
• Process: Automatically strip all personal location metadata (EXIF data) from attached images before submission; do not require a user login or any personal identifying information to submit a report.

Iterate

Iterate on the first draft through further rounds of feedback and prompting, building on what the AI tool got right and fixing what it got wrong. This will require additional prompts to instruct the AI tool what changes you want. For example, a programme officer reads the AI's first draft of a funding update, tells it which paragraph misses the point and why, and asks for a rewrite of that paragraph.

However, it is important to be aware that at some stage generative AI will start to forget context and previous instructions the longer a conversation goes on (Box X).^[bk] If a conversation runs long, check that the latest draft still matches your original brief and source material, and start a fresh conversation with that context reloaded if it does not. ^[bl]Likewise you  also need to recognise when you can’t push it any further if the outputs are getting further away from what you want. In those cases you should probably stop using AI and move on to finalising the output.

Box X. How AI forgets

An AI programme can only hold a limited amount of a conversation in view at once, called its context window. As a conversation runs longer, earlier material starts to drop out of what the AI is actively working from, even though it is still visible on your screen. A long back and forth of edits and feedback can push the source material or the brief you gave it out of view, so later drafts drift from what you originally asked for with no signal that this has happened. Where two instructions conflict, AI also leans toward the most recent one. If you asked for 5,000 words early in a conversation and later asked for 200, it will normally follow the 200 word instruction.

Verify

AI outputs can contain errors that look and sound correct. You can check any AI output by:

• Comparing it with your own sources, e.g. your data, previous reports, and personal knowledge, and with external resources, e.g. reputable news sources, published books, and journal papers.
• Asking people with relevant expertise to review it, e.g. subject specialists, local community leaders, and Indigenous Peoples, who can judge whether it is accurate and appropriate.
• Leaving out anything you have not been able to verify.

Generative AI

Generative AI can make up information and the sources for it, such as a journal paper that does not exist. It can also closely copy existing work, such as a known wildlife photograph or a published article, which can infringe the creator's copyright. To check for these problems:

• Ask the AI tool for links to the sources it used, then check them yourself.^[bm]
• Run AI generated images through a reverse image search and search for distinctive phrases from AI generated text, then credit, obtain permission for, or replace anything that closely matches existing work.

Predictive AI

A predictive model always returns its closest match from what it was trained on, even when the input does not contain enough information for a reliable answer. For example, a species identification model can confuse similar looking species, or label a species it was never trained on as one it knows. To check for these problems:

• Check that the model was trained on data similar to yours. For example, a bird call identification model trained on recordings from European woodland may perform poorly on recordings from a tropical wetland, where the species and background noise differ.
• Send outputs below a set confidence score to a person to check. For example, a forest monitoring officer sets a satellite imagery model to flag clearance automatically only when it is at least 90 percent confident, and reviews the rest by eye.
• Have a person with the right expertise review a random sample of the outputs. For example, a bat ecologist checks 100 randomly selected identifications from each month of acoustic monitoring data.
• Review every output where an error would matter most or is most likely. For example, a ranger team leader checks every area a poaching risk model marks as low risk before patrols are withdrawn from it.
• Record results only at the level of detail the input can support. For example, a survey team records grasshoppers in camera trap images at genus level, since the species can only be separated in the hand.

. 

Finalise^[bn]

You should now move on to finalise the output by:

• Removing any data, figures, or claims that verification could not confirm.
• Crediting, obtaining permission for, or replacing anything that verification found closely copies existing work.
• Correcting anything verification turned up, such as a wrong statistic or a missing citation.
• Making any other fixes to structure and content, such as reordering two sections of a report so the argument builds logically, or cutting a paragraph that repeats an earlier point.
• Disclosing AI's part in the piece whenever the audience would want to know, such as labelling an AI generated video of a simulated poaching incident used in a training presentation, in case a viewer may otherwise assume it was real footage.
• Editing for style until the piece matches the tone and format you set out to match, such as rewriting an AI drafted poster caption so it reads in the organisation's own voice.