ParallaxAI assistants are built to agree with you.

Parallax starts with one sentence, typed the way you'd say it to a friend. It doesn't answer with an essay or argue back. It asks what you mean, then what you're assuming, looks for evidence, makes the strongest case against you, and asks what changed. Beside the conversation, a map of your argument fills in, and “Where do I stand?” gives a plain audit at any point.

Parallax is the way an object shifts when you see it from two positions. A belief works the same way: you can't judge it from one place. I designed it, built the front end, and put it online at parallaxproject.co.

Result: against a chatbot told to argue with them, 6 of 8 people could name the weakest part of their own argument with Parallax, up from 2 of 8.

Chat assistants agree with people too much. Researchers call it sycophancy: a model shifts toward whatever the user seems to believe, and drops a correct answer when pushed (Sharma et al., 2023). OpenAI's own explanation of its 2025 rollback was that the update leaned too hard on the thumbs-up people give when they hear what they want.

The incentive runs the wrong way. In the Cheng study, people who talked to a flattering model rated it higher, trusted it more and wanted to use it again. Yet people find holes in someone else's argument far more easily than in their own (Mercier & Sperber, 2011), and asking them to consider the opposite reliably corrects biased judgment (Lord, Lepper & Preston, 1984). The most useful thing a conversation partner can do is disagree with you well, and that's what assistants are tuned away from.

Alongside the papers above, I took apart four products that already make disagreement useful.

I wrote one observation per sticky, one colour per person, sorted them by hand into findings, then ran a cluster pass over the sort.

• Pushback fades → five fixed stages, and dropped claims stay on the map, struck through.
• Long replies lose people (4 of 7, including the screen reader and phone users) → one short question per turn.
• Most disagreements are about a word → Clarify comes first and the opposing case fourth.
• A counter with no source reads as the machine being difficult → sources under every reply, or “Searched, nothing reliable found”.
• Views get acted on within hours → the audit is one tap away from the first reply.

In the readout, the debate coach pushed back on my plan to open with the counterargument: “Never open with the counter. Make them say what they mean first.” That moved it to stage four.

The rule everything depends on

The product never says whether you're right. It can call a claim supported, weak or contradicted. Value judgments get their own label instead of a grade. Dropped claims stay visible, struck through. The audit ends on “What would change the picture”, not a verdict.

I built in code from the start. I was designing the behaviour of a conversation, and a mockup can't show whether a question lands.

The reply is data, not only text

Each turn returns the message, the current stage, the checks run, the map nodes with types and statuses, sources and, when asked, the audit. The stepper, map and conversation never disagree because they're drawn from the same reply. The audit is its own intent, so the same question always gets the same shape of answer.

The states nobody screenshots

• Waiting: “Examining…”, announced to screen readers.
• Errors: “Try again” resends exactly what was sent, and disappears at a usage limit, where it can't help.
• Leaving mid-reply: “New belief” cancels the request in flight.
• Keyboard: Enter sends, the view switcher is a real tab list, focus rings show on glass.

I ran a within-subjects test with 8 weekly chatbot users, none from the interviews. Each brought an opinion and took it through a general chatbot told “argue against me” and through Parallax; half started with each. They did four tasks per version. I measured unaided completion and ease per task (SEQ, 1–7). The key check was whether they could name the weakest part of their own argument. The debate coach also scored each final position, blind to version, on a 1–5 reasoning rubric: 2.6 with the chatbot, 3.6 with Parallax.

• 3 of 8 found “Where do I stand?” only after scrolling the panel → moved it above the composer after the second reply → 5 of 5 found it in round two.
• 2 of 8 called Clarify stalling, and one quit there → a “Go to the opposing case” option after one clarifying answer → 5 of 5 reached the counterargument.
• The map read as a flat list on a screen reader → nested list with each status in the label → the round-two screen reader user named their weak link from the map.

Still open: one participant's ease stayed at 4; they wanted the answer faster, and Parallax is slower by design. Only 3 of 8 opened a source. And people still said the chatbot felt nicer, though 6 of 8 chose Parallax for a real decision coming up.

Don't measure whether they liked it

That last result is the measurement trap from section 02. So the north star is examined beliefs: conversations that reach Reflection or an audit. Guardrails are drop-off after the counterargument, sources that don't support their claim, and the rate of changed minds. If nearly everyone changes their view, the product is persuading, not examining.