I often know an AI-assisted draft has gone wrong before I can identify the exact sentence causing the problem.
The draft usually looks good. It has headings and transitions. The argument runs smoothly. Where there was a loose brief and an empty page a few minutes earlier, there is now something coherent enough to look nearly finished.
I have reviewed pieces like that where the surface edits were minor.
The real work was deciding that the audience needed to be narrower, that the product-first opening was solving the wrong communication problem, or that a confident benefit claim was running ahead of the evidence.
None of those decisions had much to do with sentence quality.
Generative AI makes fluency cheap. That means a draft can arrive before the question has been clarified, the evidence has been tested or the organisation has decided what it is prepared to say.
The first draft is therefore becoming a weaker measure of how far the work has actually progressed.
What is cheaper today
There is little point denying the productivity gain.
In a study involving more than 5,000 customer-support agents, access to a generative-AI assistant increased the number of issues resolved per hour by 14 per cent on average. The gains were much larger for newer and lower-performing agents. The tool appeared to help them absorb patterns that stronger colleagues had learned through experience.[1]
Knowledge work shows a similar effect.
In an experiment involving 758 consultants, participants using AI completed more work, worked faster and produced better answers on tasks that sat within the model's capabilities. On an assignment outside that boundary, however, they were less likely to reach the correct answer.[2]
That is recognisable in communications work.
Research can be summarised quickly. Structures appear almost immediately. A first pass at a landing page, article or briefing note no longer needs to consume the part of the day it once did.
Those are real gains.
But producing the first plausible version was only ever part of the work.
More of the expensive work moves into review
When production becomes easier, review can carry more of the difficult decisions.
Someone still has to decide whether the brief is asking the right question. Someone has to know which source deserves more weight, whether the argument follows from the evidence and whether an attractive sentence says more than the organisation can support.
That is where I find senior judgement entering the process most visibly.
On one run of marketing work, for example, the drafts were already clean enough that conventional copyediting was relatively minor. The senior lift came from narrowing who the work was actually for, moving from product description towards buyer value, and deciding where the available evidence was strong enough to support the intended claim.
The drafting had accelerated.
Those decisions had not.
Researchers studying what they call the “ironies of generative AI” describe a related shift from production towards evaluation. When systems make generation faster but leave people responsible for checking, correcting and integrating the output, some of the productivity gain can be absorbed elsewhere in the process.[3]
In communications work, review is rarely just proofreading.
A reviewer may need to trace a statistic to its source, separate an inference from a fact, decide whether a recommendation fits the situation, or reduce the certainty of a claim that the organisation cannot yet defend.
If the draft arrived in minutes, that judgement can become the expensive part.
Fluent output also creates pressure to make that decision quickly.
Fluency changes what the reviewer has to notice
A rough draft displays some of its uncertainty.
A polished one can conceal it.
Generative AI can place facts, estimates and assumptions in the same confident paragraph. An unresolved strategic choice can appear as settled language.
The reviewer then has to identify the uncertainty without receiving many of the visual clues that used to make incomplete thinking obvious.
A study of 319 knowledge workers found that greater confidence in generative AI was associated with less self-reported critical-thinking effort.[4] It is a self-reported measure, so I would not use it to claim that AI inevitably makes people think less.
It does point to a practical risk I recognise.
The more usable an answer looks, the easier it is to start reviewing the prose rather than the reasoning underneath it.
I pay particular attention to sentences that feel absolute before the underlying question has been settled. That is often where an assumption has disappeared into fluent language or where a recommendation is missing context that the model was never given.
The issue may only become obvious when someone with more knowledge of the business reads it.
By then, the draft can look deceptively close to done.
Better prompts still leave a decision
I use detailed briefs with AI because they improve the work.
Context matters. Constraints matter. Telling the model what is known, what is uncertain and what the output needs to achieve reduces the amount it has to infer for itself.
But a prompt is still an instruction to produce something.
It cannot know, unless someone supplies the context, that a technically correct statistic would be misleading here. It does not automatically understand why a leadership team deliberately softened a claim, or why a neat recommendation conflicts with something the organisation has already learned through experience.
The model can help me explore alternatives and expose gaps in an argument.
I still have to decide which version is sound enough to use.
That responsibility is part of the work.
Give the reviewer a defined job
NIST's generative-AI guidance treats governance, accountability, information integrity and human oversight as connected operating questions.[5]
For consequential communications, that translates into something fairly practical.
Before material leaves the organisation, I want the reviewer to be able to establish:
- which important claims come from sources that can be inspected;
- which parts are interpretation, estimate or recommendation;
- what the model supplied that nobody explicitly provided;
- what still requires verification or a decision from someone with the right context;
- who is prepared to own the final claim once it is published or sent.
AI can help with that review too.
I regularly use a model to challenge an argument, separate claims from assumptions, identify weak support and compare a draft against its source material.
That can make the checking process faster.
It does not transfer responsibility for the decision.
This is why cheaper drafting has not made experienced judgement less important in my own work. It has changed where I spend it.
Less time has to go into manufacturing the first respectable version.
More attention can go into deciding whether that version deserves to survive.
The first draft is becoming easier to obtain.
Trust is not.
Notes and sources
[1] Erik Brynjolfsson, Danielle Li and Lindsey Raymond, Generative AI at Work, NBER Working Paper 31161. Study of more than 5,000 customer-support agents. https://www.nber.org/papers/w31161
[2] Fabrizio Dell'Acqua et al., Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality, Harvard Business School / BCG. Experiment involving 758 consultants. https://aiinstitute.hbs.edu/navigating-the-jagged-technological-frontier/
[3] Microsoft Research, Ironies of Generative AI: Understanding and Mitigating Productivity Loss in Human-AI Interaction. Examines how faster generation can shift effort into verification, correction and integration. https://www.microsoft.com/en-us/research/publication/ironies-of-generative-ai-understanding-and-mitigating-productivity-loss-in-human-ai-interaction-2/
[4] Microsoft Research, The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Study of 319 knowledge workers. https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/
[5] National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1. https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence