Verna argues that models often internally score responses at low confidence, but product teams render every answer with identical typography, tone, and cadence — stripping out hedging, citations, and confidence badges. She frames this as an intentional design failure driven by demo optics, not a limitation of the underlying models.
By submitting the piece and driving it to 144 points, skadamat surfaces Verna's core claim that the UI layer — not the model layer — is where trust is being destroyed. The submission framing treats the essay as naming a pattern builders have privately noticed but not publicly challenged.
Verna names the incentive plainly: showing uncertainty makes demos land worse and loses A/B tests against confident-sounding competitors, so PMs cut hedging in review. Her implicit argument is that ethical calibration will not win on its own — the market rewards the product that always answers, regardless of whether it should.
The comment section skewed unusually unanimous for HN, with builders openly agreeing they've shipped the exact pattern Verna criticizes — proposing confidence indicators internally, watching them get cut for conversion reasons, then fielding support tickets from misled users. The consensus reaction suggests the essay is naming an open secret rather than picking a fight.
Elena Verna — former growth lead at Dropbox, Miro, and Amplitude, now at Lovable — published "Please stop the AI confidence theater" on July 2. It hit 144 on Hacker News within the day, with a comment section that skewed unusually unanimous for HN: builders agreeing they've been shipping the exact thing she's calling out.
Her argument, compressed: the dominant AI product pattern in 2026 is a chat surface that answers every question in the same authoritative tone, with the same font weight, at the same speed, regardless of whether the model is quoting a documented API or free-associating from vibes. The failure isn't that models hallucinate — it's that product teams are actively designing UIs that strip out every signal a user could use to tell hallucination from fact.
Verna calls this "confidence theater": the model may internally score a response at 40% likelihood, but the interface renders it identically to a 95% response. No hedging language, no source citations, no "low confidence" badge, no alternative interpretations offered. Just clean typography and the em-dash cadence that has become the tell of GPT-class output. She names the incentive plainly — showing uncertainty makes demos land worse, so PMs kill it in review.
This is the first widely-shared essay to name a pattern that most builders have privately noticed. Every serious AI product team has had the meeting where someone proposes a confidence indicator and someone else points out that users bounce when they see "I'm not sure." The indicator gets cut. The product ships. Two quarters later, the support queue fills with users who trusted a wrong answer because nothing in the UI told them not to.
The deeper problem Verna gestures at, but doesn't fully develop, is that calibration and conversion are in direct tension. A well-calibrated model that says "I don't know" 30% of the time will lose to a poorly-calibrated model that always answers, in every A/B test that measures engagement or task-completion-rate. The market is selecting for confident wrongness. This is not a model problem that scales away — it's a UX incentive problem that gets *worse* as models get better at sounding fluent.
Compare this to how mature engineering surfaces handle uncertainty: a compiler doesn't say "your code probably works" — it either compiles or emits a specific error at a specific line. Postgres's `EXPLAIN` shows the planner's row estimate next to the actual count so you can see when it guessed wrong. Even Google Search, for all its problems, distinguishes ads from organic results and shows you the source URL before you click. AI chat surfaces have regressed on every one of these affordances in the name of a cleaner conversational feel.
The HN thread surfaced a related pattern: RAG-based products that cite sources are quietly winning the enterprise segment while pure-chat products are winning consumer, and it's the same reason. Enterprise buyers can be fired for trusting the wrong answer, so they pay for the citation. Consumers can't, so they pay for the vibe. Verna's essay lands now because the enterprise cohort has grown large enough that "confidence theater" is finally a losing product strategy in a segment that matters. Anthropic's Claude, Perplexity, and Notion AI have all shipped inline citations in the last 18 months for exactly this reason. OpenAI's ChatGPT still doesn't, by default, in the main chat surface — which is why it still owns consumer and is losing ground in enterprise every quarter.
One comment on the HN thread cut particularly hard: "We built confidence scores into our RAG pipeline, showed them to users for two weeks, and satisfaction scores dropped 12%. We shipped without them. I'm not proud of it but I'm also not going back." That's the whole essay in one anecdote. The design problem isn't "how do we surface uncertainty" — it's "how do we surface uncertainty *without users punishing us for honesty*."
If you're shipping an AI feature in 2026, three things follow. First, instrument the gap between model confidence and user confidence. Log the model's internal probability (or a proxy — retrieval score, self-consistency across N samples) alongside whether the user acted on the answer. If users are acting on 40%-confidence outputs at the same rate as 90%-confidence outputs, your UI is lying to them and you own the consequences.
Second, stop treating "I don't know" as a failure state in your eval harness. Most eval suites score refusal as wrong. This trains the entire industry to punish the model for the one behavior that would fix the confidence-theater problem. If you're fine-tuning or doing RLHF, add an eval axis for *calibrated refusal* — the model correctly declining to answer when it doesn't have the information. A model that refuses 20% of hard questions and is right on the other 80% is more useful than one that answers 100% and is right on 70%, but only the first one shows up as "worse" in standard benchmarks.
Third, borrow from the citation-first crowd even if you're not doing RAG. Perplexity's inline `[1][2]` pattern, Claude's `
The next 12 months will separate AI products that treat uncertainty as a UX problem to solve from those that treat it as a UX problem to hide. The second cohort will keep winning consumer engagement metrics right up until the first regulatory action or class-action suit over a confidently-wrong answer that hurt someone. The first cohort will look slower and less impressive in demos and will quietly eat the enterprise market. If you're building on top of a model, the interesting question isn't which lab's weights you use — it's whether your product surface is honest about what the model actually knows. Verna's essay is the first widely-shared piece to name the pattern. It won't be the last.
My observation is that a lot of folks still discounting the capabilities or impact of AI either aren't working with frontier intelligence or aren't using it right.While the coding horse has been beat within an inch of its life already, I'd recommend throwing Codex on 5.5 high thinking
The article has solid observations, but I would correct one important thing. It's not AI confidence, it's AI psychosis.A lot of people I know are forced to use AI at work. They universally tell me that their coworkers generate awful PRs with bugs, nonsensical code and fake unit tests. But
As a hobbyist exploring LLMs since Summer 2022 (electrician by trade)... I've been as blown away by generative AI results/output as I've been left laughing at its hallucinations. My first forays were via CrAIyon & ThisWordDoesNotExist .com .Now, four years later: I've finally
> But show me something that has become so critical in your day that if I took it away tomorrow, your work would actually fall apart. Show me something truly life changing.- I no longer look at log files. If there's an MCP that gets access, I don't even download them- I don't look
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
As someone who has worked closely to the marketing space, there’s a saying that goes something like: ‘then the marketers found about it and ruin everything’. Quick example: when amazon launched kindle self publishing, there was a golden age where wannabe writers could self-publish their books, and l