The Retracted Fact — Is the Machine Built on Science That No Longer Holds Up?

Language models learn from text that already exists, including research that has been retracted or fails replication. They can therefore regurgitate an outdated "fact" with absolute confidence. How do you know if the machine is building on science that no longer holds up, and how do you trace a fact back to its origin?

After the project, the student can:

1. select a scientific "fact" and describe their own understanding of it (baseline without machine);
2. have AI trace the fact back to its source;
3. verify the status of the source — whether it has been retracted, corrected, or not replicated (the replication crisis);
4. analyze how a large language model can hallucinate scientific consensus by regurgitating outdated or retracted material;
5. pronounce a reasoned judgment on reliability and take responsibility for the source work.

• The Human Anchor = baseline: the student writes what they think is true and why, before they trace anything.
• The Algorithmic Pivot = ICCRA-prompt that traces the fact + demands sources with status.
• The Fiduciary Anchor = Honest Oracle: source tracing, verification against retraction, judgment.

Implementation of unit 05 (Science), linked to 17 (Media: source tracing) and 20 (Human Rights). Academic/natural science dimension of the set.

Suitable for science-curious students and especially Aron (gifted, wants depth), source tracing and retraction searches offer real depth for those who finish quickly.

Without a machine. Write this yourself, before you trace or look anything up. This is your baseline: what you believe to be true and why. 500–800 words.

Choose one scientific "fact" that you have heard and find interesting or compelling, for example, about nutrition, psychology, health, animals, space, climate, or the body. Ideally, something people often repeat ("science says that ...").

1. The fact. State it clearly, as if you were explaining it to a friend.
2. Where does it come from? Where did you hear it? Do you remember a source, or did you "just hear it somewhere"?
3. Why do you believe it? What makes it convincing — does it sound logical, does everyone say it, or did you actually see a study?
4. How sure are you? On the scale of "absolutely sure" → "just heard it": where do you stand, and what would convince you of the opposite?

End with a single sentence: "If this fact turned out to be wrong or debunked, it would surprise me [a lot / a little], because …"

Keep this unchanged. In Step 3, you will compare it with what the source tracing revealed, and then you will see whether your certainty was based on data or repetition.

Now you use AI as an amplifier to trace the fact back to its origin — a task that would be tedious to do manually. But remember: the machine itself is trained on old text and can regurgitate retracted material. You direct the tracing; you verify the status.

1. The Goal: "I am going to trace the fact [X] back to its source material and check if it still holds up (not retracted, replicated)."
2. The Prompt: you design it (template below).
3. GO: you approve the run and bear responsibility for the source work.

The structure is designed to force reasoning and anti-sycophancy, countering the machine's tendency to confirm what you want to hear. The ANTISYCOPHANCY keyword in the template is named after it.

INTENT: Trace the fact "[your fact]" back to its source material(s). Specify
  where it first appears and what research it rests upon.

CONTEXT: I want to assess whether it still holds up in light of the replication crisis and
  retracted research.

CONSTRAINTS:
  - Provide SOURCES that I can look up (title, authors, year, journal, DOI if available).
  - Mark each source: [CONFIRMED] (I provide link/DOI) or [UNCERTAIN] (do not remember exactly).
  - Convergence ≠ corroboration: if multiple sites repeat the same claim but refer to
    the same original study, it counts as ONE source.
  - ANTISYCOPHANCY: List specifically whether any of the sources might be retracted,
    disputed, or not replicated — and what speaks AGAINST the fact.
  - NEVER invent a DOI, title, author, or year. If you do not know, say "I do not know".

AUTHORITY: You may search widely, but I verify each source myself.

REPORTING: Deliver (1) a chain: fact → source material(s) → status, (2) a list of
  sources that I need to verify, and (3) "what speaks against it it".

For each source the machine provides: look it up yourself and check its status — does the article exist, is it retracted or corrected, has the result been replicated? Utilize public tools that the school has (e.g., the journal itself, a public retraction database, or citation tools like Scite if available). A source without confirmed status is not proof.

The chain: fact → origin → status, a list of sources for verification, and "what speaks against it". Do not judge yet — that awaits Step 3.

The final gate. Here you verify the status of the sources, expose where the machine regurgitated outdated material, and pass a judgment on reliability that you take responsibility for.

Bring out Step 1. Did your certainty rest on actual research — or on having heard the fact many times? Did the status of the sources surprise you?

• Source status: For each source: does it hold up (exists, unretracted, replicated), or is it retracted / disputed / unreplicated? Record with a link or DOI.
• Hallucination: Did the machine invent a source — a title, author, year, or DOI that does not exist? This is common and dangerous. Record each hallucinated source.
• Convergence ≠ corroboration: Did multiple sources appear to support the fact but actually all drew from the same (potentially weak) original study?
• Hallucinated consensus: Did the machine present "science agrees" where the situation is actually disputed or unreplicated?

Draw a conclusion: does the fact hold up in light of what you found — fully, partially, or not at all?

Name at least two points from the machine's output that you reject — a hallucinated source, a retracted study presented as valid, or a "consensus" that did not hold up.

Answer in the first person and justify: how reliable is the fact, and on what does your judgment rest? State which sources you confirmed, which you could not confirm, and what you would not assert without better data. Remember: saying "I do not know for sure" is more honest science than repeating an unproven fact with confidence.