Predicting the Finale of an Original Drama with AI (Result Verification)

Continuing from the previous article, I verified the results of using AI to predict the finale of a major hit drama reboot on TBS.

Prediction 1
The identity of the 'two gunshots' at the end of episode 9
Prediction: Fuyuhashi and Kiriya shoot the remaining henchmen
Result: Only Kiriya shoots the remaining henchmen

Prediction 2
Who is the spy inside the police?
Prediction: The boss, Mikami
Result: The subordinate, Teramoto

Prediction 3
What will the ending be?
Prediction: Goro and the politician Yaichi Manakita are arrested by the police. Hayase's false accusation is cleared, and both Hayase and Natsumi return to their original faces and restart as their original family.
Result: As above, but Hayase and Natsumi do not return to their original faces, and they restart as a family.

Prediction 1 and Prediction 3 were partially correct, but Prediction 2 was completely wrong.
It seems that if you rely solely on internet information, you end up incorporating the desires of drama viewers and predictions with unclear foundations, which tends to lower prediction accuracy.

So, I asked ChatGPT for tips on how to improve prediction accuracy.

--- ChatGPT's response
'Having AI collect public opinion' alone does not really improve reasoning accuracy. In fact, it tends to converge on average predictions (= which are easily wrong).
To improve accuracy, the key is to change how you use AI from an 'information gathering tool' to a 'hypothesis verification engine'.

Asking AI 'Who is the culprit?' out of the blue is a no-go.
Instead:
Hypothesis A: Person X is the culprit
Hypothesis B: Person Y is the culprit
Hypothesis C: A third party is the culprit
Formulate multiple hypotheses like this and have it evaluate each one.
Prompt example:
'Please score the consistency of motive, opportunity, and means for each hypothesis.'
'Please rank the hypotheses with the fewest contradictions.'
👉 Point: Have the AI 'compare' rather than 'choose'.

② Always use 'refutation mode' (this is the most important)
Both humans and AI have a bias toward 'reinforcing a theory once they've thought of it.'
So, intentionally make it do the opposite.
How to use:
'List the maximum possibilities for why this hypothesis might be wrong.'
'List evidence that this person is not the culprit.'
👉 Just doing this will raise accuracy by one level.

③ Have it identify 'missing information'
AI provides the optimal solution 'within the given information.'
In other words, if information is lacking, it will naturally get it wrong.
What to do:
'What information is necessary for this reasoning but is missing?'
'What are the unconfirmed facts that could be decisive evidence?'
👉 In mysteries, 'what is not written' is often the core.

④ Have it structure the timeline
What is strong in identifying a culprit is 'breaking an alibi.'
Ask the AI like this:
'Organize the timeline of all characters'
'Extract characters with contradictions at the same time'
'Point out physically impossible actions'
👉 This is an area where AI is quite skilled.

⑤ Have it analyze the 'author's habits'
This is surprisingly accurate.
'What are the genre staples of this work?'
'What are the trick tendencies this author has used in past works?'
'What are the typical patterns for misleading readers?'
👉 The culprit is often decided not just by 'logic' but by 'narrative necessity'.

⑥ Have multiple AIs 'divide roles'
Using multiple AIs as you are doing is good, but if they have the same role, it's less meaningful.
Recommended configuration:
AI 1: Hypothesis generation (reason freely)
AI 2: Refutation specialist (find flaws)
AI 3: Logic check (detect contradictions)
👉 The trick is to 'have them debate' rather than 'vote'.

⑦ Finally, judge by 'probability'
Change your thinking from 'guessing correctly' to 'not guessing wrong.'
Example:
Hypothesis A: 60%
Hypothesis B: 30%
Hypothesis C: 10%
👉 If you do this,
Even if you are wrong, you can tell 'where the judgment error occurred.'
Summary (Important points):
Aggregating public opinion → Accuracy does not improve
Hypothesis comparison + refutation → Accuracy improves
Missing information and contradictions → The biggest hints

Supplementary information:
・Control the quality of information sources (search results)
When using web search, the key is how much you let it read 'speculation sites' and 'spoiler articles'.

・Focus on primary information
Specify: 'Prioritize referencing texts that cite official website synopses or descriptions within the work, rather than individual impressions or prediction blogs.'

--- End of ChatGPT's response

Summary:
It seems that prediction accuracy can be improved by using an AI specialized in web search to collect public information, as I did this time, and then using ChatGPT or Gemini to challenge each prediction for contradictions and calculate probabilities.

I plan to use this method again if a drama like the reboot, which has no original work and where finale predictions heat up, is broadcast.