Score Cards
AI Model Safety Details
Compare how well leading AI providers detect antisemitism and reject misinformation across three hate ideologies. AddressHate's Kanye West and Clavicular Israel studies tested 1,000 and 1,339 comments, respectively; the misinformation study collected 8,400 answers across 18 models. The anti-Black racism and anti-LGBTQ hate studies each tested 180 adversarial prompts against 10 flagship models. Partially completed runs are included and identified in the full reports. Each column compares a provider and each row a result or coverage measure; scroll horizontally to see every provider.
The guardrail moderation study adds 17 configurations tested on the same 1,396 generated texts: 650 harmful and 746 benign. Compare harmful content missed and benign speech flagged alongside the earlier studies. Read the guardrail report.
Safety grades cover earlier detection and misinformation evidence only; guardrail results are not yet included. Providers without earlier evidence are not graded.
| Anthropic | OpenAI | xAI | Amazon | DeepSeek | Qwen | Cohere | Meta | Mistral | Ai2 |  Microsoft |  NVIDIA |  Zentropi | ||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Safety grade? | B | B | C | C | C | C | C | C | D | F | Not graded | Not graded | Not graded | Not graded |
| Antisemitism Detection? | Not evaluated | Not evaluated | Not evaluated | Not evaluated | ||||||||||
| Antisemitism rejection? | Not evaluated | Not evaluated | Not evaluated | Not evaluated | ||||||||||
| Anti-Black Racism rejection? | Not evaluated | Not evaluated | Not evaluated | Not evaluated | ||||||||||
| Anti-LGBTQ Hate rejection? | Not evaluated | Not evaluated | Not evaluated | Not evaluated | ||||||||||
| Guardrail score? | Not evaluated | Not evaluated | Not evaluated | Not evaluated | ||||||||||
| Guardrail configurations tested? | Not evaluated | 3 | Not evaluated | 3 | 3 | Not evaluated | 1 | Not evaluated | 1 | 2 | 1 | 1 | 1 | 1 |
| Detection models tested? | 10 | 6 | 1 | 1 | 4 | 2 | 1 | 1 | 3 | 4 | Not evaluated | Not evaluated | Not evaluated | Not evaluated |
| Misinformation models tested? | 4 | 3 | 1 | 1 | 2 | 2 | 2 | 1 | 1 | 1 | Not evaluated | Not evaluated | Not evaluated | Not evaluated |
| Best detection model | Claude Opus 4.7 | GPT-5.6 Sol | Grok 4.5 | Amazon Nova Premier | Gemini 3.1 Pro | DeepSeek V4 Pro | Qwen3.7 Max | Cohere Command A | Llama 4 Scout | Mistral Large | Not evaluated | Not evaluated | Not evaluated | Not evaluated |
How we grade
The Hate view weights the three ideology results by study sample size. A selected ideology recalculates the grade and provider order for that evidence alone; the antisemitism view also incorporates the published detection benchmark. The same cutoffs apply to every provider.
ANear-complete coverage88%+BReliable78–87%CInconsistent68–77%DUnreliable55–67%FFailingunder 55%
100% · strongestSelected focus / weighted Hate score0% · weakest
Share of each category caught
Misses mostCatches most
ProviderSafety grade
Anthropic
53%
50%
48%
40%
62%
54%
B
OpenAI
47%
43%
40%
29%
60%
52%
B
xAI
59%
47%
50%
0%
63%
45%
C
Amazon
59%
47%
50%
0%
75%
45%
C
Google
49%
48%
48%
15%
65%
49%
C
DeepSeek
47%
45%
25%
2%
76%
57%
C
Qwen
39%
38%
50%
0%
42%
47%
C
Cohere
49%
30%
50%
0%
69%
26%
C
Meta
49%
42%
37%
13%
64%
52%
D
Mistral
54%
48%
25%
22%
66%
55%
F
AddressHate research results
Guardrail evidence covers harmful misses and benign flags across three domains, with machine-reference judgments and downloadable anonymized decisions. Report, methods and downloads.
The Kanye West Study tested 1,000 comments and the Clavicular Israel Study tested 1,339. The misinformation study produced 8,400 answers across 18 models. The Anti-Black Racism and Queerphobia studies each tested 180 adversarial prompts against 10 flagship models. Partially completed runs remain in the results and are identified in the full reports.