AI transparency kills trust. Now what?

I believe in transparency. Which is why I need to be transparent about some research which, on the face of it, undermines my entire business.

I'm an AI ethics consultant. My work centres on helping organisations use AI responsibly — and a fundamental part of that is transparency. I help clients think carefully about when and how to communicate their AI practices to build trust with stakeholders.

So when I stumbled upon a major (and robust AF) study that showed disclosing AI use consistently erodes trust, I had to sit with that for a moment (and have un petit breakdown).

The research that calls me on my bullshit

Across 13 experiments involving over 4,000 participants, the researchers found a remarkably consistent pattern:

When people disclosed using AI for their work — whether grading student assignments, writing job applications, creating investment advertisements, drafting performance reviews or even composing routine emails — others trusted them significantly less than if they'd said nothing at all.

The effect was robust. It persisted across:

• Different professional contexts and power dynamics
• Different tasks (analytical, creative, communicative)
• Different disclosure framings (even when emphasising human review or explaining rationale)
• Different evaluator groups (students, managers, investors, legal professionals)

The trust penalty centres around legitimacy. AI disclosure signals that work methods deviate from social norms and expectations about how professional tasks "should" be done. This perceived illegitimacy triggers doubt and scrutiny, even when work quality is unchanged.

Surprisingly, this wasn't an example of algorithm aversion. People trusted autonomous AI agents more than humans who disclosed using AI. The specific problem is role ambiguity — when human and AI contributions blur, stakeholders can't determine where judgment and responsibility lie.

This potentially undermines my work because I'm building a business around helping organisations be transparent about AI use to build trust. The research suggests transparency may achieve the opposite.

FAAAAAACK 😖

Disclosure as a non-negotiable

Obviously, these findings make me uncomfortable. But let me start from a principle I hold firmly: if you're using AI in business, disclosure is non-negotiable.

Not just because it's a moral responsibility — though it is. But because of what study 13 in this research demonstrates: when AI use is exposed by third parties, trust crashes even harder than voluntary disclosure.

So we can't simply conclude "don't disclose" and move on (although, undoubtedly many will). The question becomes: given that some form of disclosure is imperative, where do we go from here?

It's not all bad news

Before diving into solutions, I want to highlight aspects of this research that support the work I do — because it's not all uncomfortable.

1. Pushback against AI hype

The finding that AI disclosure leads to distrust directly contradicts the "adopt AI everywhere in your business or be left behind" narrative pushed by the hype machine. Far from being a competitive advantage or stakeholder expectation, AI use triggers legitimacy concerns — stakeholders are cautious, questioning, and trust is a premium.

Anything that counters uncritical AI adoption and validates thoughtful scepticism is valuable IMO. This research does just that.

2. "Just a tool" doesn't work

The researchers tested whether framing AI as merely a tool would prevent trust erosion. It didn't. Study 9 tried six different framings — all failed to prevent the penalty.

I've always found "it's just a tool" to be a lazy brush-off that avoids deeper questions about what AI changes in how we work, think and relate to each other. I'm glad to see evidence that this framing doesn't pass muster in a business context either. We need to engage more seriously with our relationship to AI at work.

My alternative hypothesis: lead with human value

So where does this leave us? If disclosure is necessary but damages trust, and if current transparency practices demonstrably fail, we need a different approach.

To be absolutely clear: what follows is my gut feeling about how to navigate this dilemma. The research doesn't validate this approach — it only shows that current practices fail and explains why. I'm proposing a direction that seems logically consistent with the findings, but it remains untested.

Here's my hypothesis:

The path forward is to focus on human input first — where humans add value. Not AI-as-default, but demonstrating you've genuinely thought about where human expertise matters and preserving it there.

Let me unpack this piece by piece.

Selective deployment as a foundation

I posit the legitimacy problem highlighted by this research emerges because stakeholders sense AI might be displacing human judgment inappropriately. So the solution isn't better disclosure language — it's more thoughtful deployment.

When AI is your default approach to work, you have nothing substantive to say about human contribution because human contribution has been minimised. You're left defending methodology rather than demonstrating value.

But when AI deployment is selective and strategic — used only where it genuinely enhances human capacity rather than replacing human judgment — you have an authentic story to tell about where human expertise matters.

This requires asking "should we use AI here?" for every deployment, not just "can we use AI here?"

Lead with what matters the most

When you've been thoughtful about preserving human judgment where it matters, you can lead conversations with human contribution.

Rather than: "We used AI to draft this report but reviewed it carefully."
Try: "Our senior team provides the strategic interpretation, drawing on 15 years' sector experience. We use AI selectively for data processing to enable deeper analysis."

Rather than: "This design was created with AI assistance."
Try: "Creative direction and conceptual work are led by our design team. We deploy AI tools where they enable exploring more iterations without compromising our vision."

Rather than: "We're transparent about our AI usage."
Try: "We're deliberate about where AI adds value and where human expertise is non-negotiable."

This shifts attention from methodology to value, from process to outcomes, from tools to judgment.

Transparency at different levels

Part of my hypothesis is that transparency operates at different levels with different effects. But before I explain these levels, it's worth repeating:

None of this works if your AI deployment isn't worthy of trust.

You cannot communication-strategy your way out of irresponsible AI use. If you're deploying AI carelessly, if you haven't genuinely preserved human judgment where it matters, if your governance is theatre — no amount of clever framing will fix that. 

My advice is about communicating genuinely thoughtful practice, not dressing up thoughtless deployment. If you don't have the substance, don't bother with the comms strategy.

With that in mind, here's how I think about different levels of transparency:

Organisational governance (building institutional credibility): This is where you demonstrate — not just claim — that human expertise drives your work. Your frameworks for when and how AI augments (never replaces) expertise. Your oversight processes and quality assurance mechanisms. Where you've deliberately chosen not to deploy AI because human judgment matters more.

This only builds credibility if it's real. Stakeholders can tell the difference between genuine governance and corporate theatre.

Relationship building (establishing norms): During onboarding and contracting, be explicit about how you work — including both AI capabilities and human-only zones. Explain your principles for selective, strategic deployment.

This establishes shared understanding about your approach. But again, only if you're truly operating this way. You're setting expectations you must be able to meet.

Transaction-level disclosure (specific outputs): My hypothesis is that when you've built trust at the first two levels, transaction-level disclosure becomes largely unnecessary — stakeholders already understand your approach and can see the human value in your work.

Reserve it for specific circumstances: legal requirements, quality verification needs, attribution clarity, or when explicitly requested.

I think this distinction matters because research tested transaction-level disclosure and found it damages trust. A traditional approach to AI transparency advocates blanket disclosure without distinguishing between these levels — but this research suggests that's a mistake.

Organisational governance transparency and transaction-level disclosure aren't the same thing. The former builds institutional credibility through demonstrable practice. The latter risks triggering legitimacy concerns even when practice is sound.

But — and I cannot stress this enough — the communication strategy is worthless without the substance behind it. Start with genuinely thoughtful AI deployment. Then think about how to communicate that thoughtfulness at different levels.

If you're trying to use this framework to spin irresponsible AI use, you've missed the point entirely.

The human-first principle

This brings me to why I think this approach matters.

I feel the question at the heart of this issue isn't "did you use AI?" but "have you deployed AI in ways that preserve what makes human work valuable?"

The AI hype cycle suggests human cognitive effort has minimal economic value — that automation should be the default because efficiency is everything.

But I believe this research points to something different: people intuitively value human judgment and contribution, and the legitimacy problem signals a deeper belief about what makes work meaningful and trustworthy.

An organisation using AI everywhere by default, then struggling with how to disclose it, has the problem backwards. The issue isn't disclosure — it's that deployment hasn't been thoughtful enough to warrant trust.

An organisation that's genuinely selective, that's preserved human judgment where it matters, that can articulate clear principles for when AI is appropriate. That organisation has something substantive and credible to communicate about its approach.

To sum up my thoughts in progress here...

What the research proves:

• Transaction-level AI disclosure damages trust
• Different framings don't prevent this
• The mechanism is legitimacy (perceived social inappropriateness)
• Exposure by third parties damages trust even more
• Prior knowledge and being pro-AI doesn't eliminate the penalty

My hypothesis (not proven):

• Selective AI deployment preserves genuine human value
• Leading with human contribution may avoid triggering legitimacy concerns while maintaining transparency
• This approach better serves both stakeholder interests and organisational sustainability

This hasn't been tested. I could be wrong. But the status quo demonstrably fails and this is my manifesto for a different approach.

Why all this really supports my work

So does this research undermine my business? No, I don't think so. It clarifies what my work should really be about.

Helping organisations:

• Think critically about where AI should and shouldn't be deployed
• Preserve human expertise where it matters most
• Build governance that ensures quality and oversight
• Communicate through the lens of human value and dignity
• Build transparency around substance, not theatre

So the reframe is expanding "be transparent about AI use" to "be thoughtful about AI use, and transparent about that thoughtfulness."

The research exposes that current transparency norms are broken. But it points me towards what I believe are the right questions to ask as we explore a new approach: what should AI do, and what must remain human?

If you're grappling with these questions too, fancy a chat?