We made AI use something to hide

Atlassian's Teamwork Lab ran an experiment this spring in which 961 people judged the same piece of work. The only thing that changed between conditions was a line saying AI had helped produce it. The people who said so were rated about ten times lazier, and were 24 percentage points less likely to be put forward for a high-visibility project. Same work, a worse verdict for saying how it was made. Nobody needs to learn that lesson twice.

A few of the things I read last week:

• That penalty held in the same experiment even when the person framed their AI use as helping the team, which lifted the ratings a little but never back to the level of saying nothing. (Atlassian Teamwork Lab)
• Across 1,250 interview transcripts, van Nuenen, Sachdeva and Chopra find professionals managing five kinds of opacity around their AI use, and argue that universal disclosure rules will fail because what counts as inspectable depends on who is asking. (arXiv)
• In a study of 635 marketing employees who use AI, the strongest reasons for hiding it were fear of a poor evaluation, unclear policy and low trust in management. (Behavioral Sciences)
• A survey of 2,037 US workers found 90% describing themselves as confident with AI, 24.6% getting it to work on the first attempt, and 53.6% believing their senior leaders do not understand their own AI strategy. (Fortune)
• Across about 490,000 earnings-call transcripts, 95% of the sentences where firms connect AI to productivity describe gains that have not arrived yet. (St. Louis Fed)

So the people doing the work have good reasons not to say so, and the people at the top are still speaking in the future tense. Somewhere between those two we are meant to be building capability.

The silence is a sensible decision

Eric Anicich and Jeslyn Brouwers wrote in Harvard Business Review in June about a physician who had worked out prompts that genuinely helped him, and who did not pass them on to colleagues who were struggling. Their reading of the survey and interview data is that trust and psychological safety predict disclosure far better than any policy does. Mizrak and Karakaya found the same shape in their 635 marketing employees.

I do not think any of those people are behaving badly. If saying "the model drafted this" reliably costs you the interesting project, then keeping it to yourself is what a sensible person does. We built that incentive ourselves, mostly without meaning to, and then we wrote a disclosure policy and wondered why nobody used it.

The silence costs us the thing our work actually depends on. Coaching needs work you can see, and redesigning a workflow means knowing which parts of it a model is already doing. When we cannot see any of that, self-report is all we have left, which is how 90% of workers can call themselves confident while a quarter of them get the tool to work first time. We are measuring the story people tell about their work, and calling it capability.

Making it safe to say is the intervention

The same Atlassian study has a second half I found more useful than the first. In organisations where using AI is openly celebrated, the laziness penalty almost disappears, and the people who disclosed were rated as more efficient than the ones who said nothing. Nothing about the employees had changed there. What changed was what the organisation around them treated as normal.

That is something an organisation can change, cheaply, and in most places it is nobody's job. Josh Bersin noticed something related this month: HR job postings in the US are growing faster than employment overall, which he reads as HR absorbing the redesign work, because redesigning how work happens turns out to be work. As we posted on Wednesday, Gallup put the difference between teams whose manager champions AI and teams whose manager does not at 33% against 4%. Managers are where this gets settled, and they are the group we have trained the most and changed the least.

The obvious objection is that all of this leads towards surveillance, and I take it seriously. If the trouble is that we cannot see AI in the work, the tempting answer is to go and look: log the tools, mandate the disclosure, audit the drafts. Van Nuenen and his co-authors would say that fails on its own terms, because the concealment is structural rather than deceitful, and people will comply invisibly. Emily Zitek's work at Cornell suggests what that would cost. People monitored for evaluation produced fewer ideas and were readier to quit, while the same monitoring described as developmental did neither. What differed was what the looking was for. Singapore is trying a version of that at national scale, where Josephine Teo calls the aim "AI bilingualism" and the accountants' institute is putting 60,000 members through profession-run programmes instead of counting licences. I am not sure that changes how work gets organised, but I am glad someone is testing it in public.

If you work in L&D, HR, or transformation

We are usually the ones asked to raise adoption, and I think the more useful job this month is making the work visible without making it dangerous. Two things worth trying. Take one deliverable a team is proud of and ask them to walk you through where the model did the work, saying plainly that this is not an audit and none of it goes near anyone's review. Then look at what happens to the person who volunteers it anyway, and whether "I used AI for that" reads as a method or as a confession in your promotion cases and your team meetings. If nobody senior in your organisation has ever said out loud where AI helped them, that is the place to start, and it costs nothing but nerve.

The provocation

Somebody in your organisation has worked out something with AI that would help twenty other people, and has decided not to mention it. What would have to change this week for them to feel like telling you?