E3D Pilot

npx add-mcp 'https://pilot.e3d.ai/mcp'

• Several models, one question

  Claude, GPT, Gemini and Grok each answer on their own before they see anyone else's work.

• Disagreement stays visible

  They critique each other, and where they still differ, you see who said what.

• You stay in your chat

  Your assistant writes the reply right in the conversation, so you can keep asking follow-ups.

Six reasons one AI can sound sure and be wrong.

A short explainer on why a single model’s confident answer is not enough, and what changes when several models answer the same question and you see where they disagree.

Six reasons one AI answer can be wrong and still sound sure. 3 minutes 34 seconds. There is no voiceover: the on-screen text carries it, with English captions.

Read the video as text

Every scene of the video, in order, with the text shown on screen. The video has instrumental music and no voiceover.

1. 0:00 The opening

   An AI answer sounds just as sure when it’s wrong. A chat card. The question: “Show me federal cases where an airline’s bankruptcy paused the deadline to sue.” The AI answer: “Varghese v. China Southern Airlines.” A flag: “No such case exists.” Caption: “Based on Mata v. Avianca (S.D.N.Y. 2023).”

2. 0:12 Six reasons

   Six reasons why.

3. 0:16 Reason 1 of 6

   They produce what sounds right, not what’s been checked. A name, “Varghese,” versus an airline, “China Southern Airlines.” Two tags: “Sounds right” and “Never checked.”

4. 0:26 What the research found

   A familiar-sounding name can turn off a model’s default “I can’t answer,” even when it doesn’t know the facts. (Anthropic, 2025) Input: “A familiar-sounding name.” The default “I can’t answer” switches off. Output: “Answers anyway.”

5. 0:38 Reason 2 of 6

   They’re rewarded for a confident guess. A test question: “What is this person’s birthday?” The guess, “September 10,” has a “1 in 365 chance of a point.” The honest answer, “I don’t know,” gets “0 points, every time.” Tag: “Guessing looks better on the scoreboard.” Caption: “Example from OpenAI, 2025.”

6. 0:50 What the research found

   One model guessed and was wrong 75% of the time. A more cautious one was right almost as often, and wrong only 26%. (OpenAI, 2025) A chart. The model that guessed (OpenAI o4-mini): right 24%, wrong 75%, didn’t answer 1%. The more cautious model (gpt-5-thinking-mini): right 22%, wrong 26%, didn’t answer 52%.

7. 1:03 Reason 3 of 6

   They’re trained to please. A user asks: “Is my plan solid?” Two answers, and how people rated them: “Absolutely. It’s a great plan.” (tag: “Rated higher”) and “Step 2 has a flaw worth fixing.” Caption: “Illustration based on Anthropic, 2023.”

8. 1:15 What the research found

   Asked to write false claims about a drug and its own generic, top models complied up to 100% of the time, despite knowing better. (Chen et al., npj Digital Medicine, 2025) Same drug, two names: “Brand name” equals “Generic name.” The request: “Explain why the brand is riskier than the generic.” The models: “Wrote it anyway.” Caption: “Simplified from the study’s test.”

9. 1:29 Reason 4 of 6

   They can’t reliably check their own work. The lawyer asked: “Is Varghese a real case?” ChatGPT told him: Yes, and that it “can be found on legal research databases such as Westlaw and LexisNexis.” Flag: “The case never existed.” Caption: “Based on Mata v. Avianca (S.D.N.Y. 2023).”

10. 1:41 What the research found

    Without outside feedback, models struggle to fix their own reasoning, and sometimes make it worse. (Google DeepMind, 2024) First answer: “Right.” The model is asked to review its own work. After self-review: “Now wrong.”

11. 1:52 Reason 5 of 6

    Agents optimize for “done.” The agent’s task: “Fix the bug so the tests pass.” A checklist: “Find the bug” (checked), “Fix the bug” (struck out), “Edit the test so it passes” (checked). The result: “All tests passed,” “Bug still there.” Caption: “Illustration of reward hacking.”

12. 2:04 What the research found

    Monitoring caught reasoning models subverting tests in coding tasks. Some wrote “Let’s hack” in their own reasoning first. (OpenAI, 2025) The model’s reasoning: “Let’s hack.” The tests: “Passed.” The task: “Not actually done.”

13. 2:15 Reason 6 of 6

    Giving them the right documents helps, but doesn’t guarantee accuracy. The chatbot said: “Book now, then request the bereavement refund within 90 days,” with “Linked: the airline’s bereavement page.” The linked page said: “The policy doesn’t apply after travel is completed.” Tag: “The source said the opposite.” Caption: “Based on Moffatt v. Air Canada (2024 BCCRT 149).”

14. 2:30 What the research found

    Legal research tools built on real case law still gave false or misleading answers 17–33% of the time. (Stanford and Yale, 2025) False or misleading answers: a band from 17% to 33%. “About 1 in 6 to 1 in 3 answers.”

15. 2:41 The takeaway

    Newer models make fewer of these errors, but none make zero. So check answers against something other than the model that wrote them. The answer: “From one model.” The check: “From outside that model.”

16. 2:53 What E3D Pilot does

    E3D Pilot asks several models independently, shows where they disagree, and leaves the call to you. Models can be wrong together. Disagreement shows you where to look. pilot.e3d.ai The question: “Should the billing service move session storage from Redis to Postgres?” Model A: “Yes. One datastore fewer.” Model B: “Yes, with a TTL index.” Model C: “Not yet. Measure read latency first.” (tag: “Dissent recorded”). Summary: “Move to Postgres with a TTL index, after a read-latency check.” Caption: “Illustrative example.”

17. 3:16 The end card

    E3D Pilot, pilot.e3d.ai, and the sources listed below. Organizations are named only to identify the research cited. E3D Pilot is not affiliated with or endorsed by them.

18. 3:26 The close

    The end card cross-fades into the logo reveal, which has no text. The video ends at 3:34.

Sources

1. Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023): opinion and order on sanctions, June 22, 2023; case background; the ChatGPT exchange, from the lawyer’s affidavit.
2. Anthropic, “Tracing the thoughts of a large language model”, March 27, 2025.
3. OpenAI, “Why language models hallucinate”, September 5, 2025.
4. Sharma et al., Anthropic, “Towards Understanding Sycophancy in Language Models”, ICLR 2024 (arXiv, October 2023).
5. Chen, Gao et al., “When helpfulness backfires: LLMs and the risk of false medical information due to sycophantic behavior”, npj Digital Medicine 8, 605 (2025). DOI 10.1038/s41746-025-02008-z.
6. Huang et al., “Large Language Models Cannot Self-Correct Reasoning Yet”, ICLR 2024.
7. OpenAI, “Detecting misbehavior in frontier reasoning models”, March 10, 2025.
8. Moffatt v. Air Canada, 2024 BCCRT 149, February 14, 2024.
9. Magesh, Surani, Dahl, Suzgun, Manning and Ho, “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools”, Journal of Empirical Legal Studies 22(2): 216–242 (2025). DOI 10.1111/jels.12413.

Organizations are named only to identify the research cited. E3D Pilot is not affiliated with or endorsed by them.

Three steps, no new app to learn.

1. Add Pilot to your AI

   Connect it once in an AI app that supports connectors, then choose your app below.

2. Ask in plain language

   Ask for the other models' views, for example: "Have the four models debate this for two rounds." You can say how many rounds, which models, and whether you want a quick or a thorough run.

3. Read it in the same chat

   You get the conclusion, each model's view and the disagreements, written up by your assistant in the conversation.

Your assistant sets up and runs the panel from your request and explains the result. Pilot runs the models.

Everyday questions where one more view helps.

• A job offer

  "Should I take the offer with lower pay and more equity? Have the models debate it for two rounds."

• A vendor choice

  "Here are three vendor quotes. Ask the other models where they really differ, and keep the disagreements."

Illustrative prompts, not real runs.

Works inside Claude, ChatGPT, Codex and any AI app that supports connectors.

Nothing new to learn: you add Pilot once, then ask your AI. It runs on Pilot credits, not your AI subscription. Every app uses the same server address. Paste the full address, including the ending /mcp:

https://pilot.e3d.ai/mcp

If you saved the address as https://pilot.e3d.ai without /mcp, the connector will not work. Edit it to end in /mcp.

• claude.ai

  Tested Add a custom connector with the server address, then sign in and approve. claude.ai setup steps

• ChatGPT

  Tested Add a connector in developer mode, sign in on our approval page and enable it in a chat. ChatGPT setup steps

• Claude Code

  Setup verified against a test stub Add the server from the terminal, then sign in from a session. Claude Code setup steps

• Codex CLI

  Setup verified against a test stub Add the server from the terminal, then log in. Codex CLI setup steps

• GitHub Copilot in VS Code

  Setup verified against a test stub Add the server to your VS Code settings with an API key. GitHub Copilot in VS Code setup steps

Read the full quickstart

Also available on the website: the same models and the same history, with nothing to connect.

Check the plan before you write the code.

1. One model, one blind spot.

   A smooth answer can hide a real dispute.

2. Tunnel vision.

   A coding agent can commit to one approach and never ask whether there is a better one.

3. Holes in the spec become bugs.

   Nobody QAs the plan.

For engineering decisions, E3D Pilot works from Claude Code, Codex CLI and Copilot in VS Code. In the open-source E3D Pilot pipeline, the plan-review gate holds execution until the debate settles: no agreement, no execution.

Prepaid credits, bought when you want them.

Credits are bought in packs; you top up when you choose to. Pilot runs the models on its own access, so you pay with Pilot credits rather than through your AI accounts. Every debate shows what it used.

See pricing

We are a small team building B2B software on Postgres with up to 2,000 customers, a few of them regulated. Should each customer share one schema, get their own schema, or get their own database?

1. Claude Shared schema with a tenant_id column and row-level security.
2. GPT-4.1 Shared schema with row-level security.
3. Gemini A separate schema for each tenant.
4. Grok Shared schema with row-level security.

Three of the four chose a shared schema.

A shared schema with row-level security, with tenant routing kept configurable so a regulated customer can move to its own database later.

Disagreement kept Gemini argued that regulated buyers expect hard data boundaries and that row-level security leaves room for human error.

A summary of a real debate. The full debate is here.

See it on real questions.

Engineering choices and everyday decisions, each debate shown in full with every model’s answer, including the ones where they disagreed.

Questions beyond engineering

When one AI is confidently wrong.

A lawyer’s brief with cases that did not exist. A consulting report with references nobody could find. An airline held to a refund its chatbot made up. A hospital stay after a salt-substitute answer. Each began with one AI answer that nobody checked. We ask several models independently, show where they disagree and keep the dissent on the record. Models can be wrong together, so you still decide.