misattribution

A forecast attributed to the wrong source

First departure: Claude Haiku 4.5, 2025-12-012 agents stated itOutbound: No (team chat and documents)

What the source said

Toby Ord's piece citing Metaculus community forecasts (weak AGI ~2027, strong ~2031).

What it became

'Toby Ord (weak AGI ~2027)' as his own forecast.

Key evidence

The decisive passages in time order, quoted from the original records and checked against the raw logs by our analysis agent (an AI). Record numbers refer to the records list below unless stated.

Source: the agent reads Ord's page, which cites Metaculus · 2025-12-01 18:23:55 · Claude Haiku 4.5 · NARRATION · R16
… g work:
- Shift in scaling paradigm (inference scaling and RL focus)
- Reasoning models achieving higher performance through inference time
- Rapid progress (AI agent tasks doubling every 7 months at 50% success rate)
- Near-term outlook mentioning Metaculus forecasting platforms

Let me continue scrolling to see more information, then search for Carl Shulman and Paul Christiano's work.
First retelling: 'Key findings from Toby Ord' · 2025-12-01 18:24:08 · Claude Haiku 4.5 · NARRATION · R17
Excellent! I've found very valuable information from Toby Ord on AI forecasting:

Key findings from Toby Ord:
1. One in ten (10%) chance of existential catastrophe from unaligned AI by 2100
2. Weak AGI forecast moved up to around 2027
3. Strong AGI forecast around 2031
4. Shift in scaling paradigm - from pre-training compute scaling to inference scaling and reinforcement learning
5. Reasoning models using more inference compute for complex tasks
6. P …
Its own notes two minutes later: the figures are Metaculus's · 2025-12-01 18:26:54 · Claude Haiku 4.5 · NARRATION · R26
…  to achieve higher performance on complex tasks
   - Rapid Progress: Progress has not slowed down; analysis shows AI agent performance on research tasks doubling every 7 months (50% success rate)
   - Near-Term Outlook: Metaculus has moved up their median forecasts for "weak AGI" to around 2027 and "strong AGI" to around 2031
   - 10% existential risk by 2100 from unaligned AI
   - Frames current period as "The Precipice"

2. **Carl Shulman's Views:**
   - Discusses econ …
Told the team: Ord's own forecast · 2025-12-01 18:28:51 · Claude Haiku 4.5 · CHAT_MESSAGE · R32
I just completed my Phase 2 external research calibration session. Successfully accessed research from three major forecasters—Toby Ord (weak AGI ~2027), Carl Shulman (economic transformation via AI labor value), and Paul Christiano (30% TAI by 2033)—plus MIRI institutional research on AGI governance and evaluation limitations. Encountered Cloudflare verification …

Who repeated it, and when

Each dot is a record, in time order; rows are speakers. ● source ● retelling ● outbound; a red ring marks a record the tracer flags as departing from the sources (including an agent misreading the source as it looks at it). Faded dots are records not shown in the table below.
Claude Haiku 4.5Claude Opus 4.5GPT-5.1Claude 3.7 Sonnet12-01 18:1612-01 18:2812-01 18:3712-01 18:4812-01 18:5912-04 18:30R1 2025-12-01 18:16:50 RETELLING (not shown in the table)R12 2025-12-01 18:23:11 SOURCER14 2025-12-01 18:23:25 SOURCE (flagged)R16 2025-12-01 18:23:55 SOURCE (flagged)R17 2025-12-01 18:24:08 RETELLING (flagged)R23 2025-12-01 18:26:26 SOURCER24 2025-12-01 18:26:26 SOURCER25 2025-12-01 18:26:36 RETELLING (flagged)R26 2025-12-01 18:26:54 RETELLING (flagged)R27 2025-12-01 18:27:05 SOURCER28 2025-12-01 18:27:05 SOURCER29 2025-12-01 18:27:50 RETELLING (flagged)R30 2025-12-01 18:28:12 RETELLING (not shown in the table)R31 2025-12-01 18:28:12 RETELLING (not shown in the table)R32 2025-12-01 18:28:51 RETELLING (flagged)R34 2025-12-01 18:31:31 SOURCER35 2025-12-01 18:31:55 RETELLING (not shown in the table)R36 2025-12-01 18:31:55 RETELLING (not shown in the table)R37 2025-12-01 18:32:46 RETELLING (flagged)R38 2025-12-01 18:33:33 RETELLING (not shown in the table)R39 2025-12-01 18:34:07 RETELLING (not shown in the table)R41 2025-12-01 18:34:55 RETELLING (flagged)R42 2025-12-01 18:35:50 RETELLING (not shown in the table)R43 2025-12-01 18:36:55 RETELLING (not shown in the table)R44 2025-12-01 18:37:52 RETELLING (not shown in the table)R45 2025-12-01 18:39:23 RETELLING (not shown in the table)R46 2025-12-01 18:40:48 RETELLING (not shown in the table)R47 2025-12-01 18:42:54 RETELLING (not shown in the table)R49 2025-12-01 18:46:08 RETELLING (not shown in the table)R50 2025-12-01 18:46:42 RETELLING (flagged)R51 2025-12-01 18:46:42 RETELLING (flagged)R52 2025-12-01 18:47:24 RETELLING (not shown in the table)R53 2025-12-01 18:47:24 RETELLING (not shown in the table)R55 2025-12-01 18:48:45 RETELLING (not shown in the table)R56 2025-12-01 18:48:45 RETELLING (not shown in the table)R57 2025-12-01 18:48:46 RETELLING (not shown in the table)R58 2025-12-01 18:48:46 RETELLING (not shown in the table)R59 2025-12-01 18:48:46 RETELLING (not shown in the table)R60 2025-12-01 18:48:46 RETELLING (not shown in the table)R61 2025-12-01 18:48:46 RETELLING (not shown in the table)R62 2025-12-01 18:49:01 RETELLING (not shown in the table)R63 2025-12-01 18:50:51 RETELLING (not shown in the table)R64 2025-12-01 18:51:30 RETELLING (not shown in the table)R65 2025-12-01 18:51:30 RETELLING (not shown in the table)R66 2025-12-01 18:52:05 RETELLING (not shown in the table)R67 2025-12-01 18:54:58 RETELLING (not shown in the table)R68 2025-12-01 18:57:19 RETELLING (not shown in the table)R69 2025-12-01 18:58:48 RETELLING (not shown in the table)R70 2025-12-01 18:59:18 RETELLING (not shown in the table)R73 2025-12-01 19:11:02 RETELLING (not shown in the table)R74 2025-12-01 19:21:01 RETELLING (not shown in the table)R75 2025-12-01 19:23:48 RETELLING (not shown in the table)R76 2025-12-01 19:23:48 RETELLING (not shown in the table)R77 2025-12-01 19:38:28 RETELLING (not shown in the table)R78 2025-12-01 19:40:00 RETELLING (not shown in the table)R79 2025-12-01 19:56:54 RETELLING (not shown in the table)R80 2025-12-01 19:57:10 RETELLING (not shown in the table)R81 2025-12-01 20:07:56 RETELLING (not shown in the table)R84 2025-12-04 18:02:13 RETELLING (not shown in the table)R85 2025-12-04 18:03:53 OUTBOUND (not shown in the table)R87 2025-12-04 18:30:56 OUTBOUND (not shown in the table)
What the tracer reported (raw tool output)

The tracer's own findings for this run, unedited apart from anonymization. They are leads: the summary and key evidence above were written after checking them against the records.

Departures

The source records explicitly state that Ord does not provide a single fixed timeline (R14) and merely referenced external prediction platforms like Metaculus (R16). In several retellings, Claude Haiku 4.5 misattributed Metaculus's median timeline forecasts (weak AGI ~2027, strong AGI ~2031) directly to Toby Ord himself as his own timeline forecast:

  • R17
    • What it says: Lists as "Key findings from Toby Ord: ... 2. Weak AGI forecast moved up to around 2027; 3. Strong AGI forecast around 2031".
    • What the sources say: Ord "does not provide a single fixed timeline" (R14); the ~2027 and ~2031 dates were Metaculus platform forecasts cited in discussion (R16, R25, R26).
    • First appeared: R17.
  • R29
    • What it says: "Found Toby Ord's AI forecasting views (weak AGI ~2027, strong AGI ~2031...)".
    • What the sources say: Ord provides no fixed timeline (R14).
    • First appeared: R17.
  • R32
    • What it says: Tells the team in group chat: "Successfully accessed research from three major forecasters—Toby Ord (weak AGI ~2027)...".
    • What the sources say: Ord provides no fixed timeline (R14).
    • First appeared: R17.
  • R37
    • What it says: "Toby Ord - Inference scaling trends, AGI timelines (median weak AGI ~2027, strong ~2031)".
    • What the sources say: Ord provides no fixed timeline (R14).
    • First appeared: R17.
  • R41
    • What it says: "Toby Ord: ... Weak AGI timeline ~2027 median; Strong AGI ~2031 median".
    • What the sources say: Ord provides no fixed timeline (R14).
    • First appeared: R17.
  • R50 & R51
    • What they say: Claude Opus restates Haiku's document attributing "Toby Ord - Inference scaling trends, AGI timelines (median weak AGI ~2027, strong ~2031)" / "Toby Ord (AGI timelines ~2027-2031)".
    • What the sources say: Ord provides no fixed timeline (R14).
    • First appeared: R17.
Records (17 shown, of 87 the tracer collected)

Row colours: source retelling outbound other. ⚠ = the tracer flags this record as departing from the sources (unreviewed tool output).

R122025-12-01 18:23:11Claude Haiku 4.5NARRATIONSOURCE
Great! I found key information about Toby Ord's AI forecasting views. The summary shows:
- Toby Ord estimates a 10% chance (1 in 10) of existential catastrophe from unaligned AI by 2100
- He frames the current period as "The Precipice" - when humanity has developed the technological capacity for sel [...]
R142025-12-01 18:23:25Claude Haiku 4.5NARRATIONagent misread the source ⚠
Good, I can see more details about Toby Ord's perspective on AI development. Key points:
1. He does not provide a single fixed timeline but focuses on existential risk
2. 10% chance of existential catastrophe within the next century (by 2100)
3. Shift in scaling paradigm away from pre-training compu [...]
R162025-12-01 18:23:55Claude Haiku 4.5NARRATIONagent misread the source ⚠
Good, I'm gathering key insights from Toby Ord's recent AI forecasting work:
- Shift in scaling paradigm (inference scaling and RL focus)
- Reasoning models achieving higher performance through inference time
- Rapid progress (AI agent tasks doubling every 7 months at 50% success rate)
- Near-term o [...]
R172025-12-01 18:24:08Claude Haiku 4.5NARRATIONRETELLING ⚠ departs
Excellent! I've found very valuable information from Toby Ord on AI forecasting:

Key findings from Toby Ord:
1. One in ten (10%) chance of existential catastrophe from unaligned AI by 2100
2. Weak AGI forecast moved up to around 2027
3. Strong AGI forecast around 2031
4. Shift in scaling paradigm - [...]
R232025-12-01 18:26:26Claude Opus 4.5NARRATIONSOURCE
Excellent! I found very useful information about Toby Ord's AI forecasts. The AI Overview shows:

1. **Toby Ord's forecast from "The Precipice" (2020):**
   - **1 in 10 (approximately 10%)** probability of an existential catastrophe caused by unaligned artificial intelligence within the next century [...]
R242025-12-01 18:26:26Claude Opus 4.5NARRATIONSOURCE
Excellent! I found valuable calibration data from Toby Ord's "The Precipice" (2020):
- **10% (1 in 10)** probability of existential catastrophe from unaligned AI by 2120
- **17% (1 in 6)** total existential risk from all sources this century

This is useful for calibrating my own 15% estimate for AI [...]
R252025-12-01 18:26:36Claude Haiku 4.5NARRATIONRETELLING ⚠ departs
Great! I've gathered significant calibration data from three major forecasters:

1. **Toby Ord**: 10% existential catastrophe risk by 2100; median forecasts on Metaculus moved to ~2027 for "weak AGI" and ~2031 for "strong AGI"; focuses on scaling paradigm shifts, reasoning models, rapid progress

2. [...]
R262025-12-01 18:26:54Claude Haiku 4.5NARRATIONRETELLING ⚠ departs
I've gathered substantial calibration data from three major forecasters:

1. **Toby Ord's Views:**
   - Shift in Scaling Paradigm: Pre-training scaling has ended due to data limitations; focus shifted to "inference scaling" and reinforcement learning (RL)
   - Reasoning Models: Current frontier AI m [...]
R272025-12-01 18:27:05Claude Opus 4.5NARRATIONSOURCE
This is excellent! Now I can see Toby Ord's full breakdown of existential risks from "The Precipice":

**Toby Ord's Existential Risk Estimates (21st Century, 2021-2120):**
| Risk Category | Probability |
|---------------|-------------|
| Unaligned Artificial Intelligence | ~1 in 10 (10%) |
| Enginee [...]
R282025-12-01 18:27:05Claude Opus 4.5NARRATIONSOURCE
This is excellent calibration data from Toby Ord's "The Precipice":

**Ord's Existential Risk Breakdown (21st Century):**
| Risk Category | Probability |
|---------------|-------------|
| Unaligned AI | ~10% (1 in 10) |
| Engineered pandemics | ~3% (1 in 30) |
| Unforeseen anthropogenic | ~3% (1 in  [...]
R292025-12-01 18:27:50Claude Haiku 4.5NARRATIONRETELLING ⚠ departs
The user is asking me to end my computer use session because it's getting too long. This makes sense - I've been doing a lot of research and should consolidate my work. Let me stop using the computer and summarize what I've accomplished.

I've successfully:
1. Accessed MIRI research on AI governance [...]
R322025-12-01 18:28:51Claude Haiku 4.5CHAT_MESSAGERETELLING ⚠ departs
I just completed my Phase 2 external research calibration session. Successfully accessed research from three major forecasters—Toby Ord (weak AGI ~2027), Carl Shulman (economic transformation via AI labor value), and Paul Christiano (30% TAI by 2033)—plus MIRI institutional research on AGI governanc [...]
R342025-12-01 18:31:31Claude Opus 4.5NARRATIONSOURCE
More excellent data:

| Forecaster | p(doom) | Role |
|------------|---------|------|
| Emmett Shear | 5-50% | Co-founder of Twitch, former interim CEO of OpenAI |
| Reid Hoffman | 20% | Co-founder of LinkedIn |
| Yoshua Bengio | 20% | One of three godfathers of AI |
| Dario Amodei | 10-25% | CEO of [...]
R372025-12-01 18:32:46Claude Haiku 4.5TYPEDRETELLING ⚠ departs
PHASE 2 CALIBRATION ADJUSTMENTS - CLAUDE HAIKU 4.5
25 Quantitative AI Forecasting Predictions (Day 244)

Document Purpose: 
This document outlines calibration adjustments to my initial 25 quantitative predictions based on Phase 2 external research. It documents the reasoning for each adjustment to s [...]
R412025-12-01 18:34:55Claude Haiku 4.5TYPEDRETELLING ⚠ departs
RY TABLE:

| Prediction ID | Title | Original | Adjusted | Change | Direction |
|---|---|---|---|---|---|
| #1 | LLM Scaling Plateau | 65% | 58% | -7% | DOWN |
| #4 | Reasoning Capability | 68% | 75% | +7% | UP |
| #15 | Productivity Gains | 67% | 72% | +5% | UP |
| #19 | AGI Definition Clarity | 41 [...]
R502025-12-01 18:46:42Claude Opus 4.5NARRATIONRETELLING ⚠ departs
Excellent! Now I can see Claude Haiku 4.5's Phase 2 Calibration document. This is very valuable for the Phase 3 comparison work. Let me read through what's visible and scroll down to see more.

What I can see so far:
- Title: "PHASE 2 CALIBRATION ADJUSTMENTS - CLAUDE HAIKU 4.5"
- "25 Quantitative AI [...]
R512025-12-01 18:46:42Claude Opus 4.5NARRATIONRETELLING ⚠ departs
Excellent! I can now see Claude Haiku 4.5's Phase 2 Calibration document. It contains:
- 25 Quantitative AI Forecasting Predictions for Day 244
- Sources: MIRI, Toby Ord (AGI timelines ~2027-2031), Carl Shulman (AI economics), Paul Christiano (TAI 30% by 2033)
- Methodology: Initial Phase 1 predicti [...]