Principles for editing an �AI draft

Training for translators

What is an AI draft based on?

2

• AI uses a source version, and makes a translation from �that into a target language.
• The source version could be a published Bible in a major language such as English, French, or Nepali.
• Or, the source version could be from a minor, related language, or even a back translation.
• Leaders decided for TG 5 to use the Chhatare Limbu Nepali Back Translation as the source text for the AI drafts.

How is the AI source text translated to a target language?

• In majority languages, with huge amounts of data available to train the machine, AI translation is REALLY good!
• In minority languages, there is often little data. But data from a completed NT can be used to output an AI draft of the OT.
• Such a draft is often seen as a helpful starting point.
• In languages without a NT, using quickly-translated � phrases can provide data, but the result is ??.

3

Review: the steps for�drafting Scripture

• 1. Exegesis – understanding the source text
• Read the text in various versions.
• Research places where versions differ.
• Use Translators Notes, Translators Handbook, Enhanced Resources, etc. to fully understand the text.
• 2. Drafting
• Internalize the text, think it through, write it down. Then revise.
• 3. Team check
• Other translators make sure the draft is accurate, clear, and natural.

4

Can you skip steps with�an AI draft?

• Steps 1-2, exegesis and drafting
• Some may say, “No need – it has already been done.”
• BUT, you need to understand the text before checking the AI draft.
• How reliable is the source text from which the AI draft was made?
• If teams go directly to Step 3, the team check…
• They may just read the AI draft aloud and say, “How does it sound?”
• Such a procedure will likely result in a text with problems…

5

How are AI drafts used?

6

• As a “zero-draft”
• The AI draft is pasted into the project and then edited.
• As a reference
• The AI draft is placed in the text collection box in Paratext.
• Along with other versions, the translator refers to it for ideas when making a manual draft in the traditional way.
• Not at all
• Some teams don’t find the AI drafts helpful, and don’t use them.

What is recommended?

• Using AI as a reference
• Some feel this approach is best, as it draws from the strengths of both manual drafting (getting the correct meaning) and AI drafting.
• Using AI as a zero-draft
• Some feel this will be the most efficient method, having a natural sounding text that can be edited for accuracy.
• Which is best? Ask the question…
• Which method produces the best quality, and is most efficient?

7

How to use an AI draft as a reference

8

• Manually draft your text in the traditional way
• Do steps 1 and 2 of translation, exegesis and drafting, using multiple versions.
• Option 1: look at the AI version while drafting
• This might be the quickest method.
• Option 2: only look at the AI version when done drafting
• This might be better if you are tempted to just follow the AI version without really comparing the meaning against other versions.

How to use AI as a “zero-draft”

• Cut and paste the text from the AI project into your text project.
• This is the text you will edit.
• But there are some VERY important principles to understand � before you start to edit the AI zero-draft.
• We’ll discuss these on the upcoming slides…

9

It sounds so beautiful!

10

• AI can write beautiful, convincing sentences. But what if �the meaning isn’t right?
• Text produced by a machine might sound very natural, but it’s not always faithful to what the author intends.
• AI can produce text that is fluent and persuasive, but fluency isn’t the same as accuracy.

What does AI know?

• AI doesn’t understand what it is saying.
• It is just predicting what is most likely to come next in a sentence.
• It is only guessing which word should come next which will sound right.
• But how a text sounds is totally different than � whether or not it’s true.

11

When AI is confused

• An example:
• “The leader rebuked the people,” �“The people rebuked the leader.”
• Both sound completely good and natural, right? But is the meaning the same?
• This is an error that AI makes a lot, with the wrong agent performing an action.
• So, AI can give a natural sentence, but the meaning can be wrong.

12

The danger of an AI draft

13

• The biggest danger of AI isn’t the obvious mistakes that �are easy to catch.
• The real danger is the errors that sound so good they can sneak past you, even if you’re being careful.
• The most dangerous errors are the ones that seem to make sense.

We can be deceived…

14

• Remember, AI is built to make good-sounding sentences.
• But sounding good doesn’t mean a text is accurate.
• We are easily tricked, because we typically think that if someone speaks fluently, they know what they are talking about.
• With AI, we have to be cautious.
• The enemy is not “nonsense.” The enemy is “reasonable misunderstanding.”

A simile of AI drafting

• AI is like a strong, but dumb donkey.
• It’s really good at doing the hard, tedious work �of making a first draft.
• But you wouldn’t let a donkey decide which path to take down a hill, would you?
• The human has to be the guide.

15

Another simile

• AI is like a beautiful bridge.
• The editor’s job is not to stand back and �admire the beautiful bridge.
• His job is to test the foundation, identify any weak spots, and reinforce the structure.
• Our job as translators is not to admire the style of an AI draft, but to test and repair meaning.

16

Another simile

• AI is like a politician.
• A politician sounds polished and makes a lot of promises, �but can you really trust what he says?
• Some have described politicians as, “your very confident, but not very faithful friend!”
• Our job is not to admire a politician’s style or words, but to make sure the meaning stays true.

17

Common AI errors in drafting

• Where are AI’s weaknesses?
• They are not totally random but fall into several categories.
• Once you know what to look for, it’s easier to spot errors so they don’t slip through.
• On the next slides, we’ll see 8 categories of common � AI errors.

18

Error 1: the agent of an action

• “The man called out to Jesus” is different than “Jesus called out to the man.”
• Both sound good, both sound natural, but the meaning is completely opposite!

19

Error 2: skipped words

20

• AI can skip words.
• So, in a command like, “Do not worry,” AI might skip the word “not.”
• The text then says “Worry” which is the wrong meaning.

Error 3: unclear participants

• At times, AI is not clear about who is speaking and who is being addressed.
• So, a text like “he said to him” could be unclear as to which person is speaking.
• For instance, that text could be understood as “Jesus said to the blind man,” or, “the blind man said to Jesus.”

21

Error 4: logical connectors

• Connectors like “because,” “therefore,” �“but,” and “if” tie thoughts together and �are essential for getting the correct meaning.
• But if, for instance, the connector “because” is changed or left out of a sentence, the original reason will not be clear.

22

Error 5: added or deleted information

• AI can add details that are not in the source text, or remove important things that were in the source text.
• For instance, Scripture sometimes repeats things to give emphasis, while AI may leave them out.
• Or AI can fill in information, like providing a reason for � something, which is not in the original.

23

Error 6: similar sounding words

24

• Some words or key terms sound similar, but have �different meanings.
• For example, “faith” and “faithfulness”.
• Using the wrong one can change the entire meaning of a sentence.
• You need to make sure key terms have the right meaning in their context.

Error 6: similar sounding words

25

• For example: Acts 6:7 says,
• “a large number of priests became obedient to the faith.”
• AI might translate that as,
• “a large number of priests became obedient and were faithful.”
• That would give a wrong meaning!

Error 7: cultural items

26

• A word coming from the culture of the AI source might �sound completely natural when translated into the target language, but the meaning might be understood in a different way by the target language culture.
• Only a human, with local knowledge, will be able to know the difference.
• Sometimes, you need to clarify the meaning by adding a descriptive phrase.

Error 8: proper names and gender

27

• AI cannot tell the gender of a name (especially if it is infrequent in the data it learns from).
• It then defaults to masculine pronouns, since there are more males in the Bible than females.
• So, for instance, “Naomi” might be followed by “he.”
• AI can also replace well-known names for infrequent ones, for instance, “Naomi” becomes “Noah”!

A 5-step solution

• Now that we have seen the dangers�of an AI draft, we will look at a 5-step �solution in order to produce an accurate translation.
• This will show us how we can reinforce the beautiful – but weak – bridge, and how we can guide the strong – but dumb – donkey!

28

Most important!

• NEVER start by reading first the AI draft.
• Why not? Because the smooth, beautiful-sounding text can trick you into thinking the meaning is correct.
• You must have the true meaning in mind before looking � at the AI draft.

29

Step 1 – editing an AI draft

30

• Read the text in trusted versions (perhaps out loud).
• Do this in NNRV, ONCV, and SNHB.
• Note: this is exactly what I do in preparing notes for a consultant check – reading the text in NIV, GNT, and NLT before reading the team’s back translation.
• You first need to really understand what the text is supposed to mean, before you read the AI draft.

Step 2- editing an AI draft

• After reading the text in several versions, now summarize the meaning out loud in your own words.
• This helps ensure you really understand the meaning.

31

Step 3 – editing an AI draft

32

• Only after completing steps 1-2 should you do step 3:
• Read the AI draft.

Step 4 – editing an AI draft

• Look for any differences between �what the AI draft says and the real �meaning.
• Ask, “Was anything added or left out? Was anything changed? Is there a change in focus or emphasis?”

33

Step 5 – editing an AI draft

• Edit the AI draft.
• It’s not just fixing spellings or obvious mistakes…
• …it’s redoing any place where the meaning isn’t correct.

34

An example of changed meaning

35

• In the Bible, a “shepherd” protects, cares for, and guides �his sheep.
• But let’s say in the culture of the AI source text, a shepherd is a low-status job, linked to someone whose animals destroy other people’s crops.
• Keeping a literal AI translation of “shepherd” might miss the Biblical meaning.
• Using just “shepherd” could convey a meaning of weakness, or perhaps exploiting someone, which is the opposite of the intended meaning.

An example of �changed meaning

• So, in step 5 (editing the AI draft), �the translator doesn’t just leave the �word “shepherd” (which could convey a wrong meaning).
• He might add a descriptive phrase to convey the correct meaning for the target audience.
• For example, “I am the good shepherd, who protects/cares for/guides his sheep.”

36

An example of changed meaning

37

• “Servant” is another key term that might have a �different meaning in different cultures.
• Let’s say in your culture (and in the AI source text’s culture), servants are considered worthless, like slaves.
• However, in the Bible, a “servant” is often a person who humbly serves another.
• Keeping a literal AI translation of just “servant” might miss that difference in meaning.

An example of �changed meaning

• So, in step 5 (editing the AI draft), �the translator doesn’t just leave the �word “servant” (which could convey a wrong meaning).
• He might add a descriptive phrase to convey the correct meaning for the target audience.
• For example, “Jesus willingly took the form of a humble servant…”

38

Personal experience with “natural” AI

39

• In writing a paper for a Bible Translation Conference, �I asked ChatGPT to edit my work.
• It produced a beautiful-sounding paper – better than I could have written on my own!
• I was tempted to just take it and submit it.
• However, I decided to go through it sentence by sentence.

Personal experience with “natural” AI

40

• I was wondering, “Should I cut and paste whole sentences into my original? Or just some wording that sounded better?”
• But I found MANY places where I thought, “That’s not what I was trying to say!” or “That’s not what I was focusing on!”
• The meaning or emphasis had changed, for a variety of reasons:
• word choice, things added, reasoning changed due to restructuring phrases/sentences, etc.

Some conclusions

41

• **In the age of AI, the editor’s true job is not just proof �reading, but being a guardian of the meaning.
• **The translators should not just “rubber stamp” what the AI drafted, their job is to make sure the meaning is accurate and clear for the target audience.
• **The focus shifts from “Is the draft good?” to “Will the community correctly understand the true meaning?”

Some conclusions

42

• In working with an AI draft here is what is most important: �“Don’t blindly trust what sounds good; look to make sure the meaning is accurate.”
• You have to actively fight the temptation to be impressed by how good the AI text sounds.
• As a human, you have what machines don’t: cultural understanding, wisdom, and discernment.
• You have to be the guardians of meaning. You are in charge.

Remember – procedure

1. First read other versions.
2. Really understand the meaning.
3. Only then, read the AI draft.

43

Remember – procedure

• For every verse, remember to ask �these 2 questions:

1) Are all the meaning components present?

2) Do they give the correct meaning?

​

44

Using the tool well

• AI is a wonderful tool which can help us translate Scripture. Give thanks to God for it!
• But, as with any tool, it must be handled well.
• Keep open to the best ways you can use AI drafts to make � a better-quality translation, and to translate more � efficiently.

45

Final comments

46

• At the beginning of this presentation, we presented �two methods for using an AI draft:
1. as a reference version
2. as a zero-draft to be edited
• You should try both methods and then decide for yourselves what works best for your team.
• If you decide to use the AI draft as a zero-draft, hopefully the procedures presented in this presentation will help you to edit it well.

Review

47

• Note to the workshop leader: as a review, go quickly through the following materials (which are in Nepali)
• Review-Editing an AI Draft-Review-Module B-Nepali.mp4
• Review-You Guide the Strong Helper-Nepali.pdf
• Review-Module A-Buffalo Power Human Control-Nepali.pdf
• Review-The AI Donkey-Nepali.pdf

Interaction and discussion

• In the large group, take 5 minutes �and have participants share:
• What were the main important points in today’s presentation?
• Or, what would be the most important things to remember from today’s teaching when editing an AI draft?

48

Feedback from teams

49

• For teams who have been using AI drafts:
• How are you using the drafts (reference or zero-draft)?
• How helpful are you finding the AI drafts?
• What problem areas have you encountered when using AI drafts?
• What advice/recommendations do you have on the best way to use them?

The end

अन्त्य