How AI Co-Writing Shifts Writing And Opinions

A study report describes a persuasive effect of about 2–4 points on a 101-point policy-support scale.
That number can look small in isolation.
The writing process can still shift in noticeable ways.
When AI starts completing sentences, your role can move toward evaluation.
You may spend more time accepting or rejecting suggestions.
That procedural shift can affect the argument in the text.
It can also affect your opinions after writing.

TL;DR

AI co-authorship can shift writing from idea generation to evaluating suggestions, sometimes called Reactive Writing.
Small, repeatable attitude shifts have been reported, including 2–4 points on a 101-point scale.
Write a baseline position before AI output, then log why you accept suggestions and request opposing views.

Example: You draft an initial stance before enabling suggestions. You then ask the assistant to challenge your reasoning. You record why you keep or reject each suggested sentence. You later compare your final view with your initial notes.

TL;DR

What changed / what is the core issue? AI co-authorship can go beyond drafting help. It can shift users toward Reactive Writing. The loop becomes “read suggestions → accept or reject.”
Why does it matter? The change may not stop at the text. Some studies report shifts in follow-up attitude surveys. Persuasion work also reports about 2–4 points on a 101-point scale.
What should readers do? When using AI, you should write your position before AI output. You should record a reason for each adopted suggestion. For high-bias topics, you can request an opposing view and mark uncertainty.

Current state

The central question is not only “does it make me faster?”
It is also “through what procedure do I form my thoughts?”
arXiv’s Reactive Writers: How Co-Writing with AI Changes How We Engage with Ideas describes a pattern.
Reading and evaluating suggestions can become the core activity.
Accepting or rejecting suggestions can dominate the workflow.
Traditional ideation and sentence generation can receive less attention.

This procedural shift can matter beyond the text.
Co-Writing with Opinionated Language Models Affects Users' Views reports changes in expressed opinions.
It also reports shifts in follow-up attitude surveys.
Cornell University news summarizes a similar point.
It describes behavioral data like which suggestions participants adopted.
It also mentions how long it took to write a paragraph.
Those details alone do not establish key mediation.
They do not, by themselves, validate a causal pathway.

Persuasion is also not only a writing issue.
LLM-generated messages can persuade humans on policy issues reports small effects.
It reports about 2–4 points on a 101-point scale.
That effect is described as consistently small.
Small shifts may still matter in repeated workflows.

Analysis

Reactive Writing can be summarized in one sentence.
Thinking can solidify by reading and choosing among suggestions.
You may discover fewer ideas by drafting from scratch.
The AI’s first framing can become the seed of the text.
You may expand and refine that seed.

The final claim in the text can shift.
Under some conditions, post-writing attitudes may also shift.
That pathway can resemble persuasion.
It can be intentional or incidental.

A common misunderstanding focuses on persuasive sentences alone.
The procedural change can be the more central variable.
The judgment unit can shift.
It can move from “why do I believe this?”
It can move toward “should I adopt this sentence?”
Public snippets do not yet confirm quantitative causality.
That includes metrics like suggestion acceptance rate.
It also includes rewrite frequency and tone shifts.
It also includes whether evidence was sought.
Product and research teams can strengthen measurement design.
They can also improve logs for hypothesis tests.
That can help link behavior and post-attitude measures.

Practical application

From a user standpoint, “turn off AI” may not be realistic.
You can redesign the procedure instead.
The goal is not to prevent opinion change at any cost.
The goal is to control the order of opinion formation.
If AI drafts first, that draft can act as an anchor.
You can add an “AI-before” step to reduce anchoring.

Checklist for Today:

Write a short memo with your claim, reasons, and a counterargument before viewing AI output.
Add one sentence explaining why you accepted or rejected each substantive suggestion.
For sensitive topics, prompt for an opposing view and an uncertainty note before finalizing text.

FAQ

Q1. Can AI co-authorship really change my ‘opinion’?
A. Studies report shifts in expressed views and follow-up attitude surveys.
The magnitude and direction can vary across settings.
You should review each study’s conditions and limitations.

Q2. Doesn’t ‘Reactive Writing’ just mean writing becomes more convenient?
A. Convenience can be part of the experience.
The key claim is about a procedural center shift.
Reading suggestions and accepting or rejecting them can become central.
Idea generation can become less central.

Q3. At the product level, what mitigation should we test first?
A. Proposed approaches include documentation and verification prompts.
They also include opposing viewpoints and uncertainty marking.
They also include loops that confirm user intent.
Current information does not pin down direct A/B effect sizes.
A design can link behavioral logs with post-attitude measurement.

Conclusion

The impact of AI co-authorship is not only improved sentences.
It can move the center of writing from generation to evaluation.
That shift can influence what you write.
It can also influence what you believe after writing.
Product design can focus on measuring the sequence of thought formation.
It can also track post-writing opinion shifts alongside behavior logs.

Aionda

How AI Co-Writing Shifts Writing And Opinions

TL;DR

TL;DR

Current state

Analysis

Practical application

FAQ

Conclusion

Further Reading

References

Get updates