Generative AI:�What it is, how it works, �and �how to avoid the hype

Five College AI Day

May 2026

Michelle Trim

Overview

Definition of Generative AI

Generative AI is software�

Tokens and how Computers ‘Understand’ Language

Image credit: https://learn.microsoft.com/en-us/training/modules/fundamentals-generative-ai/3-language-models

How do computers make sense of natural language?

What should you take away from this explanation?

Kinds of Bias Possible in Output

• Availability bias

• Linguistic bias

• Cultural world views are conveyed in language

• Can result in incorrect translations

• Social bias

• More likely to generate stereotypical associations

• Low resourced information about underrepresented populations

• Bias in accounting of history or historical events

Errors and Confabulations – �Why might they happen

• Models are designed to deliver the next most likely token or pixel based on the training data that produced them – not to be truthful
• Errors can seem plausible, but be factually wrong, or errors can be nonsense
• Overfitting – a term that describes what happens when a model is too closely aligned with its training data such that it cannot make predictions when given new queries – it can’t make predictions that don’t appear in the training data
• More data may not always help as increasingly, text on the web is AI-generated

What GenAI can do reasonably well…

What GenAI cannot do

Generative AI cannot produce text or images that:

• Are free from Western world view bias
• Are free from biases commonly found on the internet
• Reflect an ethical thought process
• Perform qualitative evaluations or make judgements
• Name the exact text and images informing the model when it is generating a particular result
• Accurately infer emotion or emotional states

​

Do generative AI tools boost efficiency?

Measuring productivity is complex:

Software engineering - tradeoffs from decreasing manual coding include more time spent debugging, increased code review, and cognitive overloading from constant output verification.

​

Professional Tasks – study of 3500 individuals found an increase in productivity by integrating genAI tools, but, reported 20% increased boredom and an increasing loss of motivation

​

Sources:�(SadiaAfroz,ZixuanFeng,TylerMenezes,KatieKimura,BiancaTrinkenreich, Igor Steinmacher, and Anita Sarma. 2026. The Fast and Spurious: Developer Productivity with GenAI. In Companion Proceedings of the 34th ACM Symposium on the Foundations of Software Engineering (FSE ’26), June 5–9, 2026, Montreal, Canada. ACM, New York, NY, USA)

(Liu, Y., S. Wu, M. Ruan, S. Chen and X.-Y. Xie (2025, May 13). Research: Gen AI Makes People More Productive—and Less Motivated. Harvard Business Review. hbr.org)

(Rand, B. (2026, March 16). Gen AI Boosts Productivity, But Can't Turn Novices Into Experts. Harvard Business School: Working Knowledge library.hbs.edu)

(Ellis, L. (2026, Jan 21). CEOs Say AI Is Making Work More Efficient. Employees Tell a Different Story. Wall Street Journal wsj.com)

​

​

Experience and domain knowledge may be key factors:

Faster but not always better – data analysts might use AI to quickly generate marketing copy, but without marketing expertise, work is of poor quality.

​

Expectations for efficiency gains vary – C-suite executives report much higher expectations for how much time genAI is saving employees than they are reporting

​

​

​

​

​

​

​

Tasks not well aligned for GenAI tools

​

GenAI is not the right tool to bring to tasks that depend on:

• Human empathy
• Evaluation – particularly when bias is a concern
• Nuanced understanding of texts or human input
• Emotional intelligence
• Abstract criteria or instructions
• Concept mapping requiring chains of reasoning or memory

​

​

Interesting Source (https://www.mitre.org/sites/default/files/2025-03/pr-21-2414-five-ai-fails.pdf)

​

Agents

Put simply, an Agent is a software program that has the potential to act autonomously in some way. This can mean being given a high-level instruction without detailing the steps required to complete it or it can involve a complex series of tasks based on user input.

​

Agents are typically created using the following:

1. Access to specific tools or resources (accessing the web, providing specific files, IDE for modifying a code base, API for applying a translator..etc)
2. A specific goal
3. Foundation created using LLMs which provide ability to process language and act
4. Persona that establishes criteria/range of behavior desired

​

Agentic AI

Multi-agent systems rely on agent collaboration or competition to complete tasks. An example might be a system where one agent engages with a customer, another accesses a database and provides information, and another generates a return label or escalates the interaction to a human or some other personalized result.

​

Agentic AI is typically a multi-agent autonomous system acting according to predefined goals. It will collect information, formulate a strategy, undertake actions, and self correct (to an extent) to achieve its specific goal. The Canvas homework completion tool – Einstein for ex.

​

(some sources: https://www.anthropic.com/engineering/building-effective-agents; https://cloud.google.com/discover/what-are-ai-agents;)

​

Is Resistance Futile?

• Teacher agency = aligning tech engagement with larger learning goals/literacy lessons
• Student agency = building resiliency to reject, subvert, adapt genAI tools to suit one’s goals

​

In work, productivity gains may be more likely to happen when well-aligned systems are implemented in ways that augment, rather than replace, human activity.

​

Responsible Use

Data Privacy

Transparency and Disclosure

Consider Local Norms

Review Authorship Policies

Avoid Plagiarism

Remember Bias

​

Summary of Ongoing Considerations

1. Publication venues (even employers) are increasingly developing policies governing the use of generative AI for published manuscripts or work products.
2. Scholars question and have varied opinions on the ethicality and appropriateness of using AI in all research and active experiments. Students should ask an advisor or research supervisor before using an AI tool in a directed project. (Faculty - be prepared to give an answer)
3. Users are responsible for any output that is submitted in their work. If the AI gets it wrong, the user is held accountable, not the AI.
4. Models are released into the wild without large scale testing, and users may not always know that a favored tool has been updated or how an update might impact its behavior. (for ex – in 2025, OpenAI had to pull back a model update* that had made tools respond with language that was too agreeable, even harmful**)

​

*https://openai.com/index/sycophancy-in-gpt-4o/

**https://www.rollingstone.com/culture/culture-features/ai-spiritual-delusions-destroying-human-relationships-1235330175/

​

UMass Resources