GPS-bench/gps-bench-ai-bills

GPS-Bench: AI Bills and Expert Analyses

1,014 AI-governance instruments with their full text, and 429 independent expert analyses covering 259 of them, organised for leakage-controlled evaluation.

The unit of this dataset is not a bill. It is a bill paired with what human experts said about it. A model can be asked what an instrument requires, who it covers, or what it will do — and scored against analyses written by lawyers, think tanks and civil-society researchers who had no idea a model would be graded on their words.

What is here

File	Rows	Contents
gps_us_ai_bills.json	839	US federal, state and city measures
gps_international_ai_bills.json	100	National laws and frameworks outside the US, EU and China
gps_china_ai_bills.json	62	PRC laws, administrative regulations, ministry notices and standards
gps_intergov_ai_bills.json	13	4 EU instruments plus 9 from intergovernmental bodies (ASEAN, G7, OECD, UN, AU)
gps_bench_bills_WITH_analyses.json	259	The analysed subset, grouped by leakage split — the evaluation set

The four bloc files partition the corpus: every bill appears in exactly one, with no duplicate bill_key within or across them. gps_bench_bills_WITH_analyses.json is a view, not a fifth partition — it re-groups the analysed bills by split.

A file is not a bloc. gps_intergov_ai_bills.json (renamed from gps_eu_ai_bills.json on 2026-08-10) holds two: four EU instruments at bloc: "eu" and nine at bloc: "intergov". The EU legislates binding law for 27 member states; ASEAN, the G7, the OECD, the UN and the African Union issue text that binds nobody, and folding them together would erase that. In that file, read bloc off the row, never off the filename — every one of its 13 rows carries the field. Elsewhere bloc may be absent (all 62 China rows, 4 US rows and 1 international row omit it), and there the filename is the correct fallback because those files hold one bloc each.

The view is in sync with the slices. Both hold 259 bills and 429 analyses as of 2026-08-10 — the lag documented in earlier revisions of this card is closed. It remains a derived file: if you want the analysed subset under a different bar, take it from the slices with n_human_analyses >= 1 rather than from the view.

Total bill text: 47,366,541 characters. Every row carries full text: the last metadata-only row, North Carolina H301, was removed rather than kept without the statute. All 1,014 rows carry an exact YYYY-MM-DD date. bill_date_basis says which milestone that date marks — an introduction date and a promulgation date are not the same measurement, so read it before comparing dates across jurisdictions — but it is populated on only 593 of the 1,014 rows, and absent on 389 US rows (242 federal, 147 state). Where it is empty the milestone is unrecorded, not assumed.

Two date caveats, both left visible on the row rather than smoothed away. Where an AI provision was inserted into a much older statute, bill_date carries the amendment and host_statute_date keeps the original: Singapore's Parliamentary Elections Act is dated 2024-10-15 for its deepfake amendment, not 1954. Two Japanese rows have the same shape but no exact amendment date in any source, so they keep their host-statute date (1970, 2003) and carry ai_provision_year instead — filter on ai_provision_added_later to find all three. And two Georgia rows carry a future commencement date with bill_date_basis: effective; their signing dates were not collected.

US coverage is all 50 states plus the District of Columbia (214 state rows, 624 federal, 1 city). Wisconsin and Wyoming were the last two gaps and are now filled: 2023 Wisconsin Act 123 and 2025 Wisconsin Act 34, and Wyoming's 2026 Chapter 91 alongside HB 91, which failed in committee and is kept as a negative case. A state row is kept when an AI term appears in the bill's title; matching the body instead returns appropriations acts. That rule is applied consistently across the states but not across the federal subset, which was collected earlier under a looser rule — see Known gaps.

Instruments per country

41 countries, plus the EU and nine intergovernmental instruments reported separately — no member state's count absorbs the AI Act, and no country's absorbs an OECD recommendation. "Live" excludes measures that failed or were vetoed (is_live); "analysed" is rows carrying at least one expert analysis, "benchmark" those carrying two or more.

Country	Instruments	Live	Analysed	Benchmark
United States	839	611	149	76
China	62	62	35	24
United Kingdom	18	8	5	5
Canada	12	11	3	3
Intergovernmental	9	9	9	6
India	8	8	7	7
Singapore	8	8	6	6
Japan	5	5	3	3
European Union	4	3	4	4
Kenya	4	4	0	0
Saudi Arabia	4	4	4	4
Netherlands	3	3	1	0
Thailand	3	3	3	1
Ireland	2	2	2	1
South Korea	2	2	1	1
France, Russia	2 each	2 each	2 each	0
Nigeria	2	2	0	0
Australia, Brazil, Denmark, Germany, Italy, New Zealand, Qatar, Spain, Taiwan, UAE, Vietnam	1 each	1 each	1 each	1 each
Argentina, Chile, Colombia, Egypt, Indonesia, Israel, Malaysia, Mexico, Pakistan, Peru, South Africa, Turkey	1 each	1 each	1 each	0
Iran, Rwanda	1 each	1 each	0	0
Total	1,014	774	259	152

The 240 excluded are 228 US (220 federal bills that died when their Congress adjourned, plus 8 state rows — 2 vetoed, the rest failed, including Wyoming HB 91), 10 UK measures that fell at session end, the withdrawn EU AI Liability Directive, and Canada's AIDA (C-27), which died with the 44th Parliament.

Four countries — Iran, Kenya, Nigeria and Rwanda, 8 rows between them — hold bill text with no commentary at all. They are real coverage for corpus-level work and not usable as benchmark items, because a row with no analysis has no answer key. Outside the US the analysed share is high almost everywhere; inside it, 149 of 839.

Loading

The bloc files wrap their rows in a metadata object, so read the array by key rather than expecting a bare list:

import json, urllib.request

URL = "https://huggingface.co/datasets/GPS-bench/gps-bench-ai-bills/resolve/main/"

us = json.load(urllib.request.urlopen(URL + "gps_us_ai_bills.json"))["bills"]
cn = json.load(urllib.request.urlopen(URL + "gps_china_ai_bills.json"))["instruments"]

# The evaluation set is keyed by split, not wrapped in one array.
ev = json.load(urllib.request.urlopen(URL + "gps_bench_bills_WITH_analyses.json"))
after = ev["after_cutoff"]

# Filter on status_class, never on bill_status -- the latter keeps each collector's own
# words, so "is this law?" needs the normalised bucket. This works identically across
# all four blocs; matching bill_status strings does not.
enacted_us = [r for r in us if r["status_class"] == "enacted"]        # 164
enacted_cn = [r for r in cn if r["status_class"] == "enacted"]        # 29

# "Exclude failed and vetoed" is `is_live`, NOT status_class. A federal bill that died
# when its Congress adjourned still carries status_class "active", because that field
# records the furthest procedural stage reached, not the final outcome. Filtering on it
# counts 220 dead federal bills as live.
live = [r for r in us + cn if r["is_live"]]

gps_china_ai_bills.json uses instruments, not bills, because Chinese AI governance is issued as regulations, notices and standards rather than as bills passed by a legislature. The field schema is otherwise identical.

The evaluation set

259 bills carry at least one expert analysis and 152 carry two or more. gps_bench_bills_WITH_analyses.json holds exactly that set — it was rebuilt on 2026-08-10 and no longer lags the slices — and groups it by split:

Bucket	Bills	Meaning
before_cutoff	68	Earliest analysis predates the reference model's training cutoff. The memorisation baseline — a model may have read the commentary.
after_cutoff	55	Earliest analysis postdates the cutoff. Leakage-free for the reference model; the generalisation set.
train	21	Deterministic hold-out for few-shot prompting or calibration. Never score on this.
one_analysis	105	Held apart: a single analysis is real evidence but not a two-annotator item.

203 of the 259 are in force, carrying 336 of the 429 analyses; filter on is_enacted in that file for the enacted cut. Rebuild the view from the slices at any time with python3 export_merged_analyses.py.

Every benchmark-eligible row carries a split — all 152, verified. This was once a live gap: 47 rows had none, 43 having gained a second analysis after assign_splits() last ran, and 4 dropped by a builder that tested n_human_analyses != 2, an exact inequality that ejected any bill attracting a third analysis. That test now reads >= 2, and splits were assigned without re-deriving the hash hold-out, so no already-published row changed condition.

Splits are computed against claude-opus-4-5-20251101, training cutoff 2025-05-31, recorded per row in split_reference_model / split_reference_cutoff. The split is only meaningful relative to that model. For any other model, use the per-row leakage_free_for list, which names the models whose training cutoff strictly precedes the bill's earliest analysis. Models with an unpublished cutoff are omitted from that list rather than guessed at — an unknown cutoff cannot support a leakage-free claim.

Analyses per bill: 126 bills with two, 18 with three, 105 with one. A third analysis never excludes a bill from the benchmark.

Row schema

Every row carries the same core fields, in the same order, and may add optional per-row provenance after them (DOIs, archived URLs, translation blocks, recovery notes). Three invariants hold across all four slices and are worth re-checking after any edit: bill_key is unique within and across them (1,014 keys, 1,014 rows), n_human_analyses equals the number of filled human_analysis_N slots beneath it, and earliest_analysis_date is the earliest analysis_N_date actually present. Absent values are the empty string, never null.

The bill

Field	Meaning
bill_key	Stable slug, e.g. co-sb205, cn-genai-interim-2023
bill_name	Official title
jurisdiction	federal, state:XX, city:XX, eu, national:XX, international
bill_date	Exact YYYY-MM-DD on all 1,014 rows. Which milestone it marks varies by row; where bill_date_basis states it, it is not guessed.
bill_date_basis	What bill_date actually is: introduced, first_action, first_reading, first_consideration, executive_signature, chapter_assigned, promulgated, translation_published, Royal Assent and many source-specific phrasings, some a full sentence of provenance. Read this before comparing dates across jurisdictions — an introduction date and a promulgation date are not the same measurement. Set on 593 of 1,014 rows; empty on 389 US rows (242 federal, 147 state) and 32 elsewhere, where the milestone was never recorded.
bill_date_source	Which record the date came from, down to the cache file.
enacted_date / effective_date	Present where the milestone differs from bill_date and is known. Added rather than overwriting, so a row can carry both when it was introduced and when it became law.
prefiled_date / filed_date	Present on the handful of rows where the source distinguishes them from first reading.
bill_status	Procedural status exactly as its collector wrote it — 104 distinct values across the four files, and deliberately not normalised. Filter on status_class instead.
status_class	The one regime-independent bucket: enacted, active, stalled, guidance, draft, other. Every count this project publishes, including the enacted totals, comes from this field, so it is shipped rather than left to be re-derived. See below for what it does and does not claim.
bill_text	Full text of the instrument
bill_text_source / bill_text_status	Where the text came from and whether it was fetched
bill_language	Measured from the characters, not inferred from the jurisdiction
bill	Composed header + full text, for single-string prompting
bill_text_zh_primary / zh_primary	The authoritative Chinese original beside an English translation, with its provenance block (source, url, issuing body, char count). On 10 rows. PRC laws and regulations carry no copyright under Article 5 of the PRC Copyright Law, so the Chinese is freely reproducible where a translation may not be.
bill_text_en_machine / translation	A machine English rendering beside a non-English bill_text, with the model, provider, source language, chunk count and UTC timestamp that produced it. On 25 rows. Read the licensing section before using it: it is not the instrument.
bill_text_history	Present where a row's text has been restructured — what it was, what it is now, and why. Records supersessions such as a machine rendering dropped once a professional translation was found.
group	CAST catastrophe family: misuse, loss_of_control, systemic, geopolitical
group_basis	author_coded_heuristic — a hypothesis for the analyses to adjudicate, not a label
phase	1_adoption, 2_implementation, 3_societal

Progress, sponsor and outcome — added so measures can be compared across legislatures rather than only read one at a time

Field	Meaning
status_text	Where the measure got to, on one ladder: Introduced or Prefiled → Out of Committee → Passed Original Chamber → Passed Second Chamber → Sent to Executive → Enacted → Effective, plus the terminals Vetoed and Failed. This does not replace bill_status, which keeps whatever its collector said — three vocabularies, including raw govinfo version codes.
status_rank	status_text as a sortable integer. 1–7 walk the ladder, -1 is a terminal negative, 0 is unknown. The sign is load-bearing: ordering by it never claims a vetoed bill got further than an enacted one.
status_text_source	Which derivation produced it, with evidence — a tracker Status Text, a classified OpenStates action, a congress.gov latestAction, or the coded bill_status.
primary_party	Party of the primary sponsor: Democrat, Republican, Independent, Non-Designated. Non-Designated is a positive finding (nonpartisan chamber, committee bill, tracker declined to code) and is not the same claim as "", which means nobody has looked.
sponsor_party_source / sponsor_party_note	Where the party was read from, and how it was determined.
outcome_reason_kind	enacted, failed, vetoed, or "" for a measure still moving. The categorical companion to the prose in why_enacted, why_failed and outcome_detail, which cannot be filtered on.
is_live	The field to filter on for "exclude failed and vetoed." False when the final outcome was failure or a veto, or the measure is stalled or a spent draft; true otherwise. 774 of 1,014 rows are live. It exists because neither status_class nor outcome_reason_kind answers this alone — see the caution below.
key_actors	Who sponsored, backed, or is regulated by the measure.

Two tiers, because only one of them can be normalised across regimes. Whether a rule binds is a question every legal system answers, so China's in_force, the UK's Royal Assent and a US enacted all collapse to status_class == "enacted" and are safely comparable. The procedural ladder in status_text is a different matter: Passed Original Chamber → Passed Second Chamber presumes a bicameral legislature, and the NPC Standing Committee, a ministry 通知 and an EU trilogue have no honest position on it. That is why status_text is populated on the US slice and left empty on the other three — an absence by design, not a collection gap. Compare across blocs on status_class; use status_text only within a bloc.

status_class is not an outcome field, and this trips people up. It records the furthest procedural stage a measure reached, so 220 federal bills that died when their Congress adjourned still read active — their last recorded action really was "introduced / referred", and no later action exists to overwrite it. Both fields are correct as defined; they answer different questions. outcome_reason_kind gives the final outcome but is empty or absent on 596 rows, so it cannot carry a filter by itself either. is_live folds the two together and is the only field that answers "is this measure still capable of taking effect, or already law?" across all 1,014 rows. Counting live instruments by status_class == "active" overstates the US by 220.

Three cautions on status_class. guidance marks instruments that were issued but do not bind — advisory frameworks and ministry guidelines, 33 rows, held out of enacted because they are not law and out of other because they were in fact issued. And 115 federal rows reached us with a govinfo document version code in bill_status (ih (congress 118, hr), rs (congress 118, s)) in place of a status. Each such code names a pre-passage stage, so none can mean enacted and all classify as active — but that is the floor, not a recovered outcome. The status those codes overwrote is gone until the row's text is re-fetched, so treat active on a federal row as "not shown to be law", not as "still moving".

Coverage is uneven, and left visible rather than smoothed over. status_text resolves on 811 of 839 US rows; the 28 gaps are rows whose only recorded status is other or in_progress, and they are empty rather than guessed. primary_party is set on 602 of 624 federal rows but only 66 of 214 state rows, because federal sponsors come from the govinfo BILLSTATUS XML while state sponsors need the OpenStates sponsorships endpoint, which the cached sweeps never requested. key_actors, why_enacted and why_failed exist only on the 397-row curated federal subset.

The analyses — N in 1–3

Field	Meaning
human_analysis_N	Verbatim excerpt. The citable ground truth.
human_analysis_N_full	Full article text, for context
analysis_N_source / _author / _url / _date	Citation
analysis_N_source_type	law_firm_legal_analysis, think_tank, civil_society, news_media, academic_analysis, industry_analysis, other_expert_analysis
analysis_N_lean	Recorded ideological lean, where known — for controlling viewpoint, not for filtering
analysis_N_date_precision	day, month, year or none. 351 analyses carry a full date, 10 resolve only to a month, 7 only to a year and 4 have none — law-firm alerts routinely publish "December 2025" with no day, and four sources are continuously-updated trackers with no publication date at all. The partial date is kept verbatim rather than having a day invented for it. The field is not yet set on 57 of the 429 slots, all added in the 2026-08-10 rounds; absent precision is not the same as absent date.
analysis_N_date_note	Why a date is partial or absent, where it was checked.
n_human_analyses	How many slots hold a valid analysis
earliest_analysis_date / _precision	Earliest date across all three slots, and how precisely it is known. Drives the leakage split.
leakage_free_for	Model ids that could not have memorised the commentary
leakage_computable	Whether the leakage claim could be computed at all. Read this before trusting an empty leakage_free_for: an empty list means "computed, no model qualifies", which is a different claim from "could not be computed". false on 14 rows. Three of those hold analyses whose publication date is missing, leaving the earliest commentary unbounded below (la-sb-6, nyc-ll144, uk-online-safety-act-2023); the other 11 hold no analyses at all, where false is trivially true and carries no information. Treat the field as meaningful only where n_human_analyses >= 1.

How a partial date is compared to a cutoff. A leakage-free claim requires the model's cutoff to precede the analysis, so a partial date is tested at the earliest instant it could denote — 2025-12 is compared as 2025-12-01. Bounding the analysis as early as possible can only withdraw a claim, never grant one that has not been earned.

Integrity rule, inherited from CAST: no excerpt, no analysis. A row without a source name, a URL and a verbatim excerpt is not counted. Nothing in any analysis field is generated, summarised or paraphrased — every excerpt is quoted from the real article.

Who wrote the analyses

Source type	Count
law_firm_legal_analysis	170
think_tank	77
news_media	44
industry_analysis	23
academic_analysis	19
consulting_vendor	18
other_expert_analysis	15
civil_society	9
civil_society_policy_analysis	7
independent_policy_analysis	7
government_official_interpretation	5
legal_regulatory_explainer	5
Eight further types, 1–3 each	12

The vocabulary has drifted and is not yet collapsed. 20 distinct values are in use against the 7 the collection script offers, and several are one category wearing different names — academic_analysis, academic_legal_analysis, academic_expert_analysis and peer_reviewed_legal_scholarship are all academic work, 26 slots split across four labels. Group them before filtering or reporting, or a stratified sample will silently miss most of a category.

Five analyses are tagged government_official_interpretation — a ministry or regulator reading its own instrument. They are not independent commentary and are labelled so they can be excluded; among them, MIIT on its 2018 AI work plan, the CAC's technology centre on the Clear 2022 algorithm campaign, and Canada's Information Commissioner on Bill C-58, which empowers her own office.

This distribution is itself a limitation. Two in every five analyses are law-firm client alerts, a genre with a de-facto shared structure — what the law does, who is covered, obligations, effective dates, action items. Two annotators drawn from one genre agree cheaply, which is exactly what a two-annotator design is meant to test. analysis_3 exists to add a recorded, different viewpoint; analysis_N_source_type and analysis_N_lean let you stratify or restrict rather than take agreement at face value.

Languages

974 of 1,014 rows hold English in bill_text — 962 in English alone and 12 bilingual (en+zh). The remaining 22 are in 13 other languages, and 20 of them carry English beside the original:

Source language	Rows	English beside it
Spanish	6	6
Dutch	3	1
French, Thai	2 each	2 each
Danish, Indonesian, Italian, Japanese, Korean, Persian, Portuguese, Russian, Vietnamese	1 each	1 each

The two exceptions are the Dutch semiconductor rules (nl-regeling-halfgeleiders-2023 and its 2024 amendment), which hold no English side field yet. They are the only rows in the corpus a non-reader of the source language cannot work with.

No row is Chinese-only: the last such row, MOST's 2019 AI pilot-zone Guidelines, was moved onto CSET's translation, with the authoritative Chinese kept beside it in bill_text_zh_primary.

Where a row's text is not English, the original stays authoritative in bill_text and English sits beside it in bill_text_en_machine (machine translation) or bill_text_en_published (a third-party human translation), with method and provenance in a per-row translation object. A machine translation of a statute is not the statute: a mistranslated scope clause or penalty threshold reads perfectly fluently, and a model graded on "what does this provision require" would be scored against the translation rather than the law.

Three rows carry bill_text_zh_primary — the authoritative Chinese original beside a third-party English translation, because PRC laws and regulations carry no copyright under Article 5 of the PRC Copyright Law while the English renderings do.

The machine translations were produced by four different models, because the queue was split across providers to run concurrently: gpt-4.1 (6 rows), claude-sonnet-4-5 (5), deepseek-v4-pro (5), gemini-3.6-flash (3). Each row's translation.model and translation.provider record which produced it. Treat that as a caveat, not a detail: translator identity varies across rows, so an experiment that scores models on the English text is measuring translator variance alongside whatever it means to measure. For anything load-bearing, score against bill_text.

One row is a worked example of why the redundancy matters. Thailand's ETDA report was refused by one provider — two chunks returned finishReason=OTHER — and translated without incident by another. A refusal is a property of the vendor, not of the document.

Curation

Bill text comes from official sources wherever one exists: Congress.gov, OpenStates, EUR-Lex, national gazettes, cac.gov.cn, most.gov.cn. Where an official URL had rotted, the recovery is recorded on the row rather than silently substituted — the China pilot-zones instrument, for example, carries a source_recovery block noting that all twelve URLs in the source product now 404 and which mirror was used instead.

Analyses were collected by hand and by a proposal pipeline that searches a registry of publishers, fetches each hit, and verifies the bill is actually named in the article body before proposing it. A search engine matching "SB 1047" proves nothing until the text is read. The pipeline only ever proposes; a human reads the article and chooses the excerpt.

Article counts, where present, are computed by walking the numbering in document order and counting openings — not by counting regex hits, which overcounts when an article cites other articles, when an instrument cites another instrument's articles, or when a bilingual row states each article twice.

Known gaps

Stated plainly, because they bound what this dataset can support:

• 755 of 1,014 bills have no analysis at all. 690 of those are US. The corpus is much larger than the evaluation set, and most of it is bill text without commentary.
• The 852 rows with fewer than two analyses have no split, which is by design: the split is an evaluation-set property. Every one of the 152 benchmark-eligible rows now carries one. Filter on n_human_analyses >= 2 for evaluation. (bill_date was in this list until every row was dated; all 1,014 now carry an exact date. bill_date_basis, which says which milestone each marks, is still missing on 404 rows.)
• Coverage is skewed to the US (839 of 1,014) and, within it, to federal bills.
• The inclusion rule is not consistent between the federal and state subsets. A state row needs an AI term in its title, and 205 of 214 satisfy that. Only 386 of 624 federal rows do. The other 238 federal rows would fail the rule applied to the states, and because they are omnibus vehicles — NDAAs, an FAA reauthorisation, a reconciliation act — they carry 22.2M characters, about 47% of all text in the corpus, for provisions where AI is one subject among dozens. Any length-weighted statistic is dominated by them. Filter on a title match before computing one. (Those two counts come from case-insensitive matching of artificial intelligence, machine learning, algorithm, automated decision, deepfake, generative, chatbot, facial recognition, synthetic media, neural network or autonomous in bill_name, plus a case-sensitive AI token — stated so the number is reproducible rather than asserted.)
• ai_relevance is not a corpus-wide field. It is present on 744 rows — the 700 US rows that carry it and all 44 China rows — and absent from the international and intergov slices entirely, so it cannot stand in for the title filter above.
• State selection favoured enacted law. External trackers record roughly 1,200 AI bills introduced across the states in 2025 and about 145 enacted; this corpus holds a small share of the former and most of the latter. It is close to a census of enacted state AI law and a thin, non-random sample of what was merely introduced. Do not read state counts here as a measure of legislative activity.
• Many bill_date values on China rows are translation dates, not issue dates. Rows sourced from CSET carry bill_date_basis: translation_published, and the gap runs to years: the 2022 scenario-innovation Guiding Opinions carry a late-2025 date because that is when the English appeared. split derives from bill_date, so on those rows it describes the translation's timing rather than the instrument's. Check bill_date_basis and original_publication_date before treating bill_date as the date Chinese actors first saw the policy.
• One row (nyc-ll144) has two analyses whose source pages carry no publication date. It counts toward the two-analysis total but contributes nothing to the leakage split, and its leakage_free_for is deliberately empty. Two further analysed rows are leakage_computable: false for the same reason — la-sb-6 and, in its third slot, uk-online-safety-act-2023.
• Two Dutch rows carry no English translation — nl-regeling-halfgeleiders-2023 and nl-regeling-halfgeleiders-2024-wijziging. Every other non-English row has English beside the original.
• coverage_area is empty on 63 rows. It comes from a semantic classifier and was left blank rather than guessed on rows the classifier never ran over.
• The (Trial) Measures for AI Technology Ethics Review row runs to Article 37 but opens only 36 — the position where Article 9 should stand reads Article 4 in the source PDF. Recorded, not silently corrected.

Licensing — read before redistributing

This dataset mixes three rights regimes and is not uniformly redistributable.

1. Bill text is generally not copyrightable. US federal and state legislation, EU instruments, and PRC laws and regulations (Article 5, PRC Copyright Law) are public texts.
2. human_analysis_N_full contains 429 whole third-party articles from 302 distinct sources — 4,544,965 characters of law-firm client alerts, think-tank commentary and trade-press writing. These are copyrighted works reproduced in full. They are included for research reproducibility. Redistribution, commercial use, or training on this field may not be covered by fair use or fair dealing in your jurisdiction. The verbatim human_analysis_N excerpts are short quotations and sit on much firmer ground than the full texts.
3. Six China rows hold English pasted from ChinaLawTranslate, whose robots.txt sets Content-Signal: search=yes,ai-train=no,use=reference. Five predate this collection; cn-ai-labeling-draft-2024 was added on 2026-08-11 the same way, by paste rather than fetch, because that signal is honoured rather than reopened — a programmatic fetch of that site succeeds, which is exactly why declining it has to be deliberate. Two carry the public-domain Chinese original in bill_text_zh_primary; four do not. Per-row terms are in bill_text_access_note. If you are assembling training data, exclude these six rows.
4. bill_text_en_machine is model output, not a translation of record. 25 rows carry one, produced by translate_non_english_bills.py and labelled on the row with the model and provider that made it. It exists so a non-English instrument can be searched and triaged in English. It is not the instrument, it has not been checked against the original, and bill_text remains authoritative on every one of those rows. Do not score a model against it and do not promote it into bill_text.

If you need a cleanly licensed corpus, drop human_analysis_N_full and keep human_analysis_N plus analysis_N_url — the excerpt and the citation.

Intended use, and what this is not

Built for evaluating LLM reasoning about AI governance: what an instrument requires, who it binds, how it interacts with neighbouring law, and what experts judged its effect would be. The leakage split is the point — scoring only on before_cutoff measures recall of published commentary, not analysis.

This is not legal advice, not a complete census of AI legislation, and not a gold-standard annotation set. The analyses are expert commentary, not adjudicated truth; two experts frequently disagree, and group is an author-coded hypothesis the analyses are meant to test rather than a validated label.

Citation

@misc{gpsbench_ai_bills,
  title  = {GPS-Bench: AI Bills and Expert Analyses},
  author = {Melanie},
  year   = {2026},
  note   = {Dataset of AI-governance instruments paired with independent expert analyses,
            organised for leakage-controlled evaluation},
  url    = {https://huggingface.co/datasets/GPS-bench/gps-bench-ai-bills}
}

Please also cite the original publishers of any analysis you quote — each row carries analysis_N_source, analysis_N_author and analysis_N_url.

Changelog

2026-08-11 — China slice expanded, and a translation layer

• Corpus 1,014 instruments (was 996), 429 analyses on 259 bills (was 411 on 249), 152 benchmark-eligible (was 144). The growth is entirely the China slice: 44 → 62.
• Eighteen PRC instruments added against the Sinograph governance-document register, via a new add_china_bills.py — the China slice had no adder, and its rows had been placed by hand. Among them the four foundational laws the corpus had been missing (Cybersecurity Law as amended 2025, Data Security Law, PIPL, Law on Progress of Science and Technology), both TC260 AI Safety Governance Frameworks, and China's four outward-facing governance documents (Global AI Governance Initiative and Action Plan, Shanghai Declaration, AI Capacity-Building Action Plan).
• Six rows now carry bill_text_en_machine. Six PRC instruments have no published English translation, so translate_non_english_bills.py attached a labelled Gemini rendering beside the Chinese. bill_text was not touched on any of them. See licensing item 4 before using that field.
• Ten rows now carry bill_text_zh_primary — the authoritative Chinese sitting beside an English translation rather than behind it.
• Two documents on the register are recorded as unavailable rather than approximated: TC260-005 is served from a session-gated endpoint, and the AI Application Safety Classification and Grading method has no document at all — its page is a call for drafting units. Seven GB standards remain out because std.samr.gov.cn publishes metadata, not standard text.

2026-08-10 (b) — card correction

The data files were current; this card was not. Every count below was re-measured against the four slices and the numbers throughout the card now match them.

• gps_eu_ai_bills.json no longer exists. It was renamed gps_intergov_ai_bills.json and grew from 4 rows to 13, taking on nine intergovernmental instruments (ASEAN, G7, OECD, UN, African Union) alongside the four EU ones. The eu config in this card's header still pointed at the old filename, so the dataset viewer had a config resolving to a file that was not there; it is now intergov, pointing at the file that is. If you loaded this dataset with load_dataset(..., "eu"), use "intergov". The bloc field on each row is what separates EU from intergovernmental — read it off the row, not off the filename.
• Corpus 996 instruments (was reported as 984), 411 analyses on 249 bills (was reported as 398 on 241).
• gps_bench_bills_WITH_analyses.json has been rebuilt and no longer lags. It holds all 249 analysed bills and 411 analyses. The previous card described a deliberate 209-bill freeze; that freeze is over, and any result computed against the 209-bill version should be re-derived or pinned to a Hub revision before 2026-08-10.
• Corrections to claims that were wrong rather than merely stale: bill_date_basis is populated on 592 of 996 rows, not all of them; guidance is 33 rows, not 5; bill_status holds 104 distinct values, not 71; outcome_reason_kind is empty on 596 rows, not 537; leakage_computable: false appears on 14 rows, 11 of which hold no analyses at all and where the flag carries no information.
• New known gaps recorded: two Dutch rows carry no English translation, so the earlier "every non-English row carries English beside it" no longer holds; ai_relevance is absent from the international and intergov slices; coverage_area is empty on 63 rows.
• Removed references to count_by_country.py, validate_slices.py and schema/gps_slice.schema.json, none of which ship with this dataset. The invariants they were said to check are stated directly instead, and all three hold.

2026-08-10 (a)

• Corpus grew from 952 instruments, and analyses from 356. Almost all of the growth is international: that slice went from 54 to 100 rows and from 14 countries to 41, adding the Gulf (Saudi Arabia, UAE, Qatar, Egypt), Latin America (Chile, Colombia, Peru, Argentina, Mexico), and Indonesia, Iran, Israel, Malaysia, Pakistan, South Africa, Taiwan, Thailand, Turkey.
• 20 non-English rows gained bill_text_en_machine or bill_text_en_published; bill_text is untouched and every bill_text_sha256 still validates. The machine translations come from four different models — see Languages for why that is a caveat and not a footnote.
• Dating is no longer a known gap: all 996 rows carry an exact YYYY-MM-DD date.
• New caveat recorded rather than fixed: analysis_N_source_type has drifted to 20 distinct values against the 7 the collection script offers, with academic work split across four labels. Group before stratifying.

2026-08-08

• Corpus grew from 922 to 952 records. Eleven came from the international ingest — Kenya (4), Nigeria (2), Rwanda (1) and four UK measures — and nineteen were recovered from the local benchmark build, which had rows present in no published slice. The recovered rows are tagged merged_from and carry fewer fields than their neighbours.
• Added the normative layer to every slice: normative_character (hard_law / soft_law / non_instrument) and the coverage_area block with its classifier basis, confidence and runner-up. coverage_area is the sector an instrument governs, never its geography.
• Ten UK measures are now is_live: false with outcome_reason_kind: failed_session_end — they fell when their legislative session ended rather than being defeated on the merits.
• All four slices were brought to one key order and passed the project's internal validator (validate_slices.py against schema/gps_slice.schema.json — build tooling that does not ship with this dataset). Ten fields the card already documented were missing from that schema, so nothing validated them; declaring them is what caught a boolean in an object field and an out-of-enum status value.
• Known gap: the nineteen recovered rows have no coverage_area yet. It comes from a semantic classifier and was left empty rather than guessed.