Disclosure Is the Feature

Why telling people it is AI now beats hiding it.

For two years the smart play in consumer Artificial Intelligence (AI) was to hide it. Bury the model behind a clean interface, skip the label, keep sign-up friction low. That play has expired. This week a United States bill moved to protect the data people type into chatbots, the European Union locked in its content-labelling deadline, and fresh bias findings showed what opacity costs when a system gets a hiring or policing call wrong. The through-line across all five stories is the same: trust in everyday AI is now built by disclosure, not concealment. Issue 018 maps the inversion and shows you how to act on it before it becomes the minimum everyone meets.

1. Congress moves to ban the sale of health and location data typed into AI chatbots

United States lawmakers want to close an Artificial Intelligence gap in privacy law. Senator Elizabeth Warren and Representative Mary Gay Scanlon are preparing a version of the Health and Location Data Protection Act that would bar the sale of health and location data to brokers, and it names the data people enter into chatbots such as ChatGPT and Claude as explicitly in scope. The bill is still at proposal stage, so it is a signal rather than a statute. The signal is the point. The sensitive information consumers hand to everyday AI, their symptoms, locations and worries, is moving from a product-design question to a compliance one. It arrived the same week the United Kingdom's Information Commissioner's Office (ICO) brought new data-protection complaint duties into force, so the direction is not confined to Washington.

Read more: Rep. Scanlon (official) and 9to5Mac

McGann's Take: Consumers are starting to price in what happens to what they type. Brands that say plainly what they collect and delete will read as honest; the ones that stay quiet will read like the brokers this bill targets.

2. AI accountability week: hiring-tool bias and quietly shelved UK predictive policing

Three findings in one week sharpened the cost of opaque Artificial Intelligence. Stanford's Institute for Human-Centered AI (HAI) published what it called the first large-scale study of hiring algorithms tested in the wild, documenting racial bias and systemic candidate rejection. In England, a WIRED investigation found that Avon and Somerset police had built more than twenty predictive risk models, two of which were quietly shelved over genuinely poor performance. Separately, tests of leading chatbots reported measurable political lean at odds with vendors' neutrality claims. Read together, the message is that "the model works on average" no longer settles anything when the system decides who gets hired, policed or advised. Accountability is moving out of ethics panels and into operational and legal exposure.

Read more: Stanford HAI

McGann's Take: Opacity used to be a shortcut. Now it is the liability. If you cannot show how a customer-facing model reached its answer, assume a regulator or a claimant eventually will.

3. Anthropic ships Claude Tag, an always-on AI teammate

Anthropic launched Claude Tag, a persistent enterprise agent that replaces its previous Slack app with a standing teammate that learns a company's context, keeps memory across tasks, and acts on its own rather than answering turn by turn. It fits a clear direction of travel: agents as always-on infrastructure rather than a chat window you open and close. The same week, MIT Technology Review cautioned that dressing agents up as digital coworkers can quietly erode human oversight, because people stop checking a colleague they trust. That tension is the story. As everyday AI shifts from something you prompt to something that runs in the background, whether people know an agent is acting, and on what authority, stops being cosmetic.

Read more: VentureBeat

McGann's Take: An agent working quietly in the background is convenient right up to the moment it is wrong. Telling people where it operates is not friction; it is the audit trail.

4. Anthropic accuses Alibaba of an industrial-scale model "distillation" attack

Trust in Artificial Intelligence is not only about what firms tell consumers; it is about whether the model is even what it claims to be. Anthropic told the United States Congress that Alibaba had extracted its Claude model's capabilities through nearly 29 million exchanges run from fraudulent accounts, an industrial-scale "distillation attack" meant to copy a stronger system into a weaker one. Alibaba denies wrongdoing and has separately sued to be removed from a Pentagon blacklist, so the claims remain contested. Set the dispute aside and the point still stands: provenance is becoming a first-order question. If a capable model can be cloned at scale, buyers and consumers need to know whose system they are actually using and where its answers come from. That is the same instinct now driving disclosure rules, applied one layer up the supply chain.

Read more: BBC News

McGann's Take: Provenance is the new trust primitive. Consumers will not audit training data, but they will remember which brands were straight about what sits under the hood.

5. Europe finalises its AI Act transparency code, live from 2 August

Europe has turned disclosure from good manners into law. The European Commission published the final Code of Practice on transparency for AI-generated content under Article 50 of the European Union (EU) AI Act, ahead of the obligations taking effect on 2 August 2026. In plain terms, systems that generate or manipulate content will have to mark it as machine-made, and providers will have to make that signalling workable in practice. It lands against a tense backdrop, with Europe negotiating for access to United States models from a position of some weakness. The compliance clock matters, but the strategic shift matters more. Once labelling is mandatory in your largest adjacent market, hiding the AI stops being a design choice and becomes a legal risk.

Read more: Tech Policy Press

McGann's Take: August is a deadline, not a strategy. The winners will treat mandatory labelling as a chance to look trustworthy, not a box to shrink to the smallest legal font.

Disclosure Is the Feature

Here is the specific claim. In 2026, for consumer-facing Artificial Intelligence, telling people a system is AI now improves adoption more than hiding it does. That is a reversal. Until roughly late 2025 the opposite held: hiding the model kept friction low and sign-ups high, so the smart money concealed it. The inversion is not a mood. It is structural, and three forces drove it.

First, non-disclosure has become operationally fragile. The European Union's content-labelling obligations take effect on 2 August 2026. A United States bill is moving to govern the data people type into chatbots. The United Kingdom's Information Commissioner's Office has new complaint duties in force, and the Financial Conduct Authority (FCA) now reports that more than 80% of financial services firms already use AI, which is precisely why supervisors are turning to labelling and provenance. When your biggest markets require a signal, concealment stops being clever and starts being exposure.

Second, the cost of hiding has become legible. This week alone produced documented racial bias in hiring tools and predictive-policing models withdrawn for poor performance. Once harm is attributable, "we would rather not say how it works" reads as a decision, not a neutral default. Customers, journalists and regulators now know to ask.

Third, disclosure is starting to build preference, not merely satisfy compliance. Provenance disputes, such as Anthropic's complaint over model distillation, show that even sophisticated buyers care whose system they are using. Ordinary consumers will not read a model card, but they remember which brands were straight with them. Early, confident disclosure signals a company with nothing to hide, and that reads as quality.

Hiding the AI was a 2024 strategy. Disclosing it is a 2026 strategy.

Name the strongest objection, because it is real. In some categories, disclosure measurably lowers engagement. Tell readers a piece was AI-assisted and some stop reading; label a recommendation as algorithmic and click-through can dip. In creative writing and content recommendation, concealment still looks better on a weekly dashboard. But that is a short-horizon read. The engagement you keep by hiding is borrowed against the day a customer finds out, and discovery is now more likely, not less, given labelling laws and competitors who disclose with confidence. A dip in clicks is recoverable. The moment a user feels deceived is not. Disclosure costs something today and pays back over a longer horizon than most dashboards measure.

So do not treat disclosure as a legal chore to shrink. Take your highest-traffic customer touchpoint that uses AI, a support chat, a recommendation, an onboarding flow, and run the disclosure audit in this issue's Prompt Pocket on it this week. Write two versions of the disclosure: one bare, and one that adds the single piece of context that would reassure a wary user, such as human review or a data-deletion control. Measure trust, not only clicks. The brands that move first will own "clearly AI" as a mark of quality before it becomes the minimum everyone has to meet.

Why this tool fits the audience: In a week about everyday consumer trust, Pi is built for the exact question this issue asks: where do ordinary people feel AI is on their side? Inflection AI tuned Pi for emotional intelligence and personal use rather than work throughput, which makes it a fair test of warmth, privacy clarity and value for non-technical users, not raw productivity.

Pricing: Free for personal use at time of writing (£0 / $0). Pi has no consumer paid tier; enterprise and Conversational Application Programming Interface (API) access is custom-quote only, so there is no public figure to convert. As at 01 Jul 2026. Consumer AI pricing moves quickly, so verify on pi.ai within four days of send.

Try it: pi.ai →

This is the editorial argument made executable: a disclosure audit you run on one real customer experience. Paste in a description of a single customer-facing thing that uses AI behind it, a chatbot, a recommendation, a generated message, an automated decision, and the prompt tells you what to disclose, how to phrase it, where to place it, and the one piece of context that would lift trust most. Use it before you ship a new AI feature, or on an existing one you have never labelled. Tactical note on reading the output: treat the risk score as the trigger and the "lift" line as the real work; the phrasing is easy to change, but the context you choose to add is what a wary user actually feels. Run it in Claude or ChatGPT as written.

You are a trust-first product strategist. I will paste a description of one
customer-facing experience that uses AI behind the scenes (a chatbot, a
recommendation, a generated message, or an automated decision).

For that experience, return five things:

1. DISCLOSE: exactly what should be disclosed, the fact that AI is involved,
   the process, or both. Be specific to this experience.
2. PHRASE: two ways to word the disclosure, one short label and one fuller
   sentence that names the benefit to the user.
3. PLACE: where the disclosure should sit, a visible badge, an in-flow line
   before the interaction, or a note after, and why.
4. LIFT: the single piece of extra context that would raise trust most here
   (for example human review, a data-deletion control, or an accuracy caveat).
5. RISK: score the reputational risk of NOT disclosing, 1 to 10, and say why.

Number your answer. End with the one question I must answer before I can
implement this.

The question that prompt hands back, which single piece of context would most reassure a wary user, is where the real decision lives, because it forces you to name the thing you have been hoping customers would not ask.

The inversion is not a forecast; it already happened this quarter. What is still open is how fast brands act on it, so here are two predictions you can hold me to.

By 30 June 2027, at least one major UK consumer brand will make "clearly disclosed AI" a headline in a marketing campaign, not a line buried in a privacy policy. And by 31 December 2026, the share of UK consumers who actively prefer disclosed AI over hidden AI in the tools they use daily will cross 60%, measurable through YouGov or a comparable tracker. If either misses, I will say so here.

The brands that win the next year will not be the ones with the best model. They will be the ones whose customers believe them. Disclosure is how you earn that, and the work starts with one honest sentence on one screen.

If you ran the disclosure audit on one customer experience this week, hit reply and tell me the one line you would change first, and what it currently hides. I read every response.

John McGann

Founder, Zymbos AI