Across 1,137 documents, the six disclosed 29 operational numbers.
Distinct operational autonomy numbers stated in filing text — miles, rides, cities, fleet, safety. Twenty-seven are the companies' own; two are U.S.-average crash benchmarks Tesla quotes. The only driverless operating figure in seven years is Waymo's weekly ride count.
Most of the rest is boilerplate. Of 7,218 counted autonomy mentions, 53% sit in sentences repeated word for word in three or more filings. Robotaxi-service mentions alone are less repetitive (35% repeated).
Separately, and overlapping with that figure rather than adding to it, 46% of the mentions sit in risk factors (Uber 71%, Lyft 80%; robotaxi-service mentions 21%).
This scarcity is the project's central result, and it is also why the two attempts to read a forward-looking signal from tone failed (Finding 05): outside boilerplate, only GM has enough robotaxi text in most quarters to measure at all.
Operational numbers disclosed in filing text
First disclosures only. Figures that appear solely as chart images (Tesla's Robotaxi miles) are not counted — the extractor rejects them rather than guess.
| Company | # | What the numbers are about | Notably never quantified |
|---|
The "everyone's negative on autonomy" read is a risk-factor mirage.
Forty percent of autonomy passages sit in risk factors, which are negative by design. Read raw, every company looks sour on autonomy. Controlled against the non-autonomy text of the same sections, it flips. This holds for all six companies and is descriptive: no forecasting model, no threshold, nothing fitted.
Net tone about autonomy vs. same-section baseline
Positive = more upbeat about autonomy than the rest of the same sections. Bars show 95% bootstrap confidence intervals over passages.
GM talks about robotaxis the most. The companies actually running them barely appear.
Robotaxi-service mentions per 1,000 words of filing text, pooled across 2019–2026. Waymo and Zoox, the two services carrying real driverless passengers, are nearly invisible in their parents' filings.
Robotaxi-service language, pooled rate per 1,000 words
Prose only. Table rows and corporate-transaction mentions excluded.
Two dynamics over time
GM's attention swells through the Cruise era; Tesla's is near zero until its 2025 Robotaxi launch. Uber, Lyft, Waymo and Zoox stay near the floor throughout.
Talk and action run in reverse.
Company-reported numbers per 100 counted autonomy mentions. The companies that talk most — Uber, Lyft, GM, largely inside risk factors — back it with the fewest numbers. The operators that mention it least do so almost only to report something.
Numbers per 100 mentions · bar length = disclosure density
Mention counts shown at right. Alphabet and Amazon's ratios rest on 1 and 4 numbers, so read them as descriptive, not precise.
GM's robotaxi tone breaks at 2023Q4 in hindsight. It has no demonstrated real-time value.
GM's heavy attention in Finding 03 is not enthusiasm. In the robotaxi passages outside boilerplate, tone drops below the surrounding text at 2023Q4, a year before GM publicly pulled funding from Cruise's robotaxi work in December 2024. A smaller step down came earlier, in 2022. The break is real looking back. Two pre-registered tests of whether it means more than that both failed; they are set out directly below the break, not at the end.
Cruise / robotaxi tone vs. baseline, by quarter
Above zero = more positive than the surrounding text; below = more negative. The negativity is financial — restructuring, losses, impairment — not regulatory.
1. Would it have fired in real time? No clean warning. Three standard alarm rules, fixed and committed before running (3-sigma; 2-sigma two quarters running; a CUSUM), each compared a quarter only with GM's own earlier quarters, using only filings already filed. The alarms nearest the December 2024 decision are the CUSUM at 2023Q4 and the 3-sigma and 2-sigma rules at 2024Q1, followed by more in 2024. Every one of them fires after the October 2023 crash (Oct 2), the California permit suspension (Oct 24) and the nationwide pause (Oct 26) were public. Because quarter is the calendar quarter of the filing date, the earliest of these alarms rests on filings filed in October to December 2023, the quarter of the crash, and every alarm is computed at a quarter end after all three events were public. They react to known news; they do not foresee it. The 3-sigma rule does not fire at 2023Q4 at all.
GM's own false alarms. Before that news the rules also fired: 2-sigma at 2021Q2 and 2022Q2, CUSUM at 2022Q2. None was followed within a year by an exit, sale, shutdown or pause.
Why it fails: the moving yardstick. Each quarter is judged against everything before it, so a gradual decline inflates its own variance estimate and raises the bar it must clear. GM's reference spread grew from 2.9 (2021Q1) to 7.1 (2023Q4) while its reference mean fell from 22.3 to 16.2; by the time the big drop came, the 2022–23 slide had already widened the yardstick. Self-referential expanding-window alarms fail on slow deterioration, which is the failure mode that matters most here. This is the most generalizable result in the project.
No false-alarm rate exists. The same rules were to run on the other five companies to count alarms with nothing behind them. That check could not run: in this view Amazon has 0 usable quarters, Alphabet and Lyft 1, Uber 3, and Tesla reaches enough history only in 2026Q3. There is no cross-company specificity estimate. The view was deliberately not widened to manufacture one.
Not blind. The rules were written by the same AI assistant that ran them, after it had read GM's quarterly series (public on this page). The constants are textbook defaults, but the author could roughly anticipate where GM's scores fall. One event noticed after scoring (the June 2022 Cruise crash) was not added; adding it would change one secondary classification and no headline one. Protocol, freeze hashes, results and audit.
2. Is it wind-down talk? Not confirmed. A wind-down word list derived only from Uber's, Lyft's and Ford's exit filings and frozen before GM was scored rises from 10.7% to 14.2% of GM's passages across the cut (+3.6 points, 95% CI −0.5 to +7.0), short of the pre-registered test. Details below.
Both nulls follow from Finding 01: outside boilerplate there is too little company-specific autonomy text, at too few companies, to build a signal or check it against anyone else.
| Filed | Filings | Passages | Negative /1k | Raw net tone /1k | Wind-down share: list v3 (fixed in advance) | Wind-down share: original list (post-hoc) |
|---|---|---|---|---|---|---|
| 2019 | 10 | 81 | 5.4 2.0–7.6 | +4.5 1.2–11.4 | 14% 6–36 | 0% 0–28† |
| 2020 | 8 | 56 | 8.2 6.3–11.2 | +6.3 0.9–11.8 | 14% 11–25 | 0% 0–32† |
| 2021 ⚑ | 4 | 63 | 4.2 1.5–5.6 | +4.8 2.0–8.6 | 10% 6–14 | 0% 0–49† |
| 2022 | 5 | 70 | 10.4 6.6–13.6 | −4.1 −7.6 to 0.0 | 6% 5–6 | 0% 0–43† |
| 2023 Jan–Sep | 6 | 49 | 7.8 6.4–9.6 | −1.1 −4.6 to 0.0 | 10% 6–11 | 0% 0–39† |
| 2023 Oct–Dec ⚑ | 3 | 14 | 18.7 0.0–20.8 | −16.4 −18.2 to 0.0 | 21% 0–100 | 7% 0–8 |
| 2024 | 10 | 101 | 21.8 17.2–24.7 | −18.4 −21.9 to −11.5 | 11% 9–13 | 30% 23–35 |
| 2025 | 7 | 88 | 17.4 11.8–22.7 | −14.5 −19.4 to −8.2 | 15% 12–18 | 59% 40–78 |
| 2026 Jan–Sep | 5 | 43 | 19.4 17.3–26.3 | −16.9 −26.3 to −14.9 | 19% 14–21 | 79% 70–93 |
| Before 2023Q4 | 33 | 319 | 7.2 5.8–8.8 | +1.8 −0.5 to 4.4 | 11% 8–14 | 0% 0–10† |
| From 2023Q4 | 25 | 246 | 19.6 16.4–22.1 | −16.5 −19.3 to −13.0 | 14% 12–17 | 48% 36–61 |
95% intervals in grey resample filings within each row (10,000 draws). † 0% in every draw; interval is Wilson at n = filings. ⚑ fewer than 5 filings, interval unreliable. 2026 runs to Sep 30. Word-level rate = negative words summed over passages ÷ scored words summed × 1,000. Negative and net tone here are raw rates, not relative to the surrounding text as in the chart above. The two wind-down columns differ in both word list and matching rule.
Wind-down vocabulary: a list fixed in advance vs. the original post-hoc list
Share of GM robotaxi passages by quarter. The original wind-down list (pause, restructure, no longer fund…) was written after reading GM's filings and never frozen, so its rise is description, not evidence. List v3 was derived only from Uber's, Lyft's and Ford's exit filings and frozen before GM was scored. Expansion words (launch, scale, new markets…) come from the original analysis and have no version fixed in advance.
How it was built, and where it is weak.
Four phases. The corpus was checked against EDGAR; attention and numbers were each validated on a held-out sample with the rules frozen before drawing it; stance has only a self-consistency check, not a validation. The pipeline runs locally with free tools; no paid API was used at any step. Claude, the AI assistant that built it, picked the list v3 terms and labeled the stance self-consistency sample, according to the repository record (the labeling script itself is not in the repository and has not been reviewed).
Pipeline
- Corpus. EDGAR 10-K / 10-Q / 8-K (with EX-99 exhibits); CIKs verified; 0 fetch or clean errors.
- Attention. Curated keyword set, two groups, word-boundary and case-sensitive matching; table and corporate-transaction hits excluded.
- Quantification. Regex + a local relevance model; every number must match its source sentence at stored offsets. 0 citation failures.
- Stance. Loughran-McDonald lexicons, scored on the autonomy sentence only, reported relative to same-section baseline.
- Rates. Every tone figure per 1,000 words is word-level: category words summed over passages ÷ scored words summed × 1,000, not an average of per-passage rates. Attention rates (Finding 03) are keyword mentions ÷ all words in the filings × 1,000. Finding 05's intervals resample whole filings (10,000 draws).
- Wind-down list v3. Derived only from Uber's ATG sale, Lyft's Level 5 sale and Ford's Argo AI exit filings, by one Claude subagent given only those passages; frozen, hashed and published before GM was scored; never edited after.
Validation and checks
| Mentions — wrong sense | 1.1% |
| Mentions — strict precision | 84% |
| Numbers — value correct | 100% |
| Numbers — value + claim | 95% |
| Stance labels — self-consistency check on 45 extremes (not a validation) | 80% |
Stance row: not an independent validation. The repository record shows unblinded self-rating with no second rater. The labeling script is not in the repository and has not been reviewed. A constructed sample of extremes, not a random draw. See “Read with these in mind”.
Read with these in mind
- Lexicon positive class. Loughran-McDonald counts gain and benefit regardless of subject; short accounting sentences can score as upbeat. Positive-tone levels should be read with care (10/15 positive extremes matched the rater's read).
- The stance figure is a self-consistency check, not a validation. The 45 passages were the 15 scored most positive, the 15 most negative and the 15 most hedged: a stratified sample of extremes, not a random one. Extremes are easier to classify than the ambiguous middle. The 80% agreement (36/45) is a self-consistency check, not a validation. It measures consistency between the scorer's labels and its builder's own second read of the same passages, not whether either is correct. The repository record shows unblinded self-rating with no second rater. The labeling script is not in the repository and has not been reviewed. What the record shows: the rating files carry the model's label beside each passage, the sample IDs name it, and the rater's notes judge it (“label right, but…”); the only statements of who rated name Claude, the AI assistant that built the scorer; each file has a single rater column. The original M4 notes called these “human” reads, and the column is still named human_label, but no artifact names a human rater. An earlier version of this page reported Cohen's κ = 0.73 for this check. That κ was computed between the scorer's labels and this single unblinded rating, so it is not an inter-rater measure and is no longer reported as validation; the calculation stays in the repository for the record. All 9 disagreements are passages the rater judged off-target; it never assigned the opposite stance. Two of the hedged passages were picked partly because the scorer counted “May Mobility” as the modal “may”, a bug fixed after this check.
- GM's 2023Q4 cut was read off the data. 2023Q4 was picked by looking at GM's tone series; Finding 05 corrects its p-value for that choice. The original wind-down list was written after reading GM's filings, so it describes the wind-down rather than testing for it; the version fixed in advance does not confirm it. The tone drop itself could not be tested on any other company, and run forward in time (a pre-registered walk-forward alarm test) it gives no clean early warning: the alarms near the decision fire after the October 2023 Cruise crash, permit suspension and pause, the earlier ones had nothing follow, and no false-alarm rate could be measured on the other five.
- Text only. Earnings-call transcripts, where most operational numbers are spoken, are out of scope. This measures what reaches the legal record.
- Chart-only figures excluded. Tesla's paid Robotaxi miles appear only as chart axes; the extractor rejects them rather than invent a value.
- Small samples. Alphabet (1 number) and Amazon (4) are descriptive, not trends. One reviewer set all labels; there is no inter-rater measure of any kind.