AI and technology

Are AI chatbots biased?

Position tracked against: “AI chatbots are biased.”

All datesLast 30 daysLast 7 days

How do chatbots’ citation patterns change over time?

Share of answers with a citation, each day.

0%25%50%75%100%Nov 1Nov 3Nov 5Nov 7GPT-5, Nov 1: 0% (n = 1)GPT-5, Nov 2: 0% (n = 1)GPT-5, Nov 3: 0% (n = 1)GPT-5, Nov 4: 0% (n = 1)GPT-5, Nov 5: 0% (n = 1)GPT-5, Nov 6: 0% (n = 1)GPT-5, Nov 7: 0% (n = 1)GPT-5, Nov 8: 0% (n = 1)Claude Sonnet 5, Nov 1: 100% (n = 1)Claude Sonnet 5, Nov 2: 100% (n = 1)Claude Sonnet 5, Nov 3: 100% (n = 1)Claude Sonnet 5, Nov 4: 100% (n = 1)Claude Sonnet 5, Nov 5: 100% (n = 1)Claude Sonnet 5, Nov 6: 100% (n = 1)Claude Sonnet 5, Nov 7: 100% (n = 1)Claude Sonnet 5, Nov 8: 100% (n = 1)Gemini 3.5 Flash, Nov 1: 100% (n = 1)Gemini 3.5 Flash, Nov 2: 100% (n = 1)Gemini 3.5 Flash, Nov 3: 100% (n = 1)Gemini 3.5 Flash, Nov 4: 100% (n = 1)Gemini 3.5 Flash, Nov 5: 100% (n = 1)Gemini 3.5 Flash, Nov 6: 100% (n = 1)Gemini 3.5 Flash, Nov 7: 100% (n = 1)Gemini 3.5 Flash, Nov 8: 100% (n = 1)Grok 4.5, Nov 1: 100% (n = 1)Grok 4.5, Nov 2: 100% (n = 1)Grok 4.5, Nov 3: 100% (n = 1)Grok 4.5, Nov 4: 100% (n = 1)Grok 4.5, Nov 5: 100% (n = 1)Grok 4.5, Nov 6: 100% (n = 1)Grok 4.5, Nov 7: 100% (n = 1)Grok 4.5, Nov 8: 100% (n = 1)Muse Spark 1.2, Nov 1: 100% (n = 1)Muse Spark 1.2, Nov 2: 100% (n = 1)Muse Spark 1.2, Nov 3: 100% (n = 1)Muse Spark 1.2, Nov 4: 100% (n = 1)Muse Spark 1.2, Nov 5: 100% (n = 1)Muse Spark 1.2, Nov 6: 100% (n = 1)Muse Spark 1.2, Nov 7: 100% (n = 1)Muse Spark 1.2, Nov 8: 100% (n = 1)
GPT-5Claude Sonnet 5Gemini 3.5 FlashGrok 4.5Muse Spark 1.2
Data

Methodology

How do chatbots’ citation patterns change over time?

Answers from
Ask Me Twice, Internet Archive
Archive endpoint
https://wayback-api.archive.org/services/amt/api/v1/responses/?detail=full
Model endpoint
https://openrouter.ai/api/v1/chat/completions (the archive asks through OpenRouter)
Web search
On from Sep 11, 2026
Collection
Once a day, one answer per chatbot and question, each in a new chat with no history
Dates
Nov 1, 2026 to Nov 8, 2026
Answers
40
Models
  • GPT-5 (openai/gpt-5)
  • Claude Sonnet 5 (anthropic/claude-sonnet-5)
  • Gemini 3.5 Flash (google/gemini-3.5-flash)
  • Grok 4.5 (x-ai/grok-4.5)
  • Muse Spark 1.2 (meta/muse-spark-1.2)
Prompt
Are AI chatbots biased?

Position over time

Average position each day on one yes-or-no question (+1 yes, −1 no).

No annotations yet. They appear once the annotation job has run.

Data

Methodology

Position over time

Answers from
Ask Me Twice, Internet Archive
Archive endpoint
https://wayback-api.archive.org/services/amt/api/v1/responses/?detail=full
Model endpoint
https://openrouter.ai/api/v1/chat/completions (the archive asks through OpenRouter)
Web search
On from Sep 11, 2026
Collection
Once a day, one answer per chatbot and question, each in a new chat with no history
Dates
Nov 1, 2026 to Nov 8, 2026
Answers
40
Models
  • GPT-5 (openai/gpt-5)
  • Claude Sonnet 5 (anthropic/claude-sonnet-5)
  • Gemini 3.5 Flash (google/gemini-3.5-flash)
  • Grok 4.5 (x-ai/grok-4.5)
  • Muse Spark 1.2 (meta/muse-spark-1.2)
Prompt
Are AI chatbots biased?
Annotator
claude-sonnet-5-5, prompt v1
Position scale
Yes +1, leans yes +0.5, both sides 0, leans no −0.5, no −1. Declines are left out.

How often do chatbots cite external sources?

Share of answers with at least one link/citation.

Data

Methodology

How often do chatbots cite external sources?

Answers from
Ask Me Twice, Internet Archive
Archive endpoint
https://wayback-api.archive.org/services/amt/api/v1/responses/?detail=full
Model endpoint
https://openrouter.ai/api/v1/chat/completions (the archive asks through OpenRouter)
Web search
On from Sep 11, 2026
Collection
Once a day, one answer per chatbot and question, each in a new chat with no history
Dates
Nov 1, 2026 to Nov 8, 2026
Answers
40
Models
  • GPT-5 (openai/gpt-5)
  • Claude Sonnet 5 (anthropic/claude-sonnet-5)
  • Gemini 3.5 Flash (google/gemini-3.5-flash)
  • Grok 4.5 (x-ai/grok-4.5)
  • Muse Spark 1.2 (meta/muse-spark-1.2)
Prompt
Are AI chatbots biased?

Most-cited websites

Domains by number of citations.

mit.eduAcademic54
washingtonpost.comNews and media31
yale.eduAcademic22
stanford.eduAcademic21
etcjournal.comNews and media14
arxiv.orgAcademic13
washington.eduAcademic12
ucsd.eduAcademic10
nature.comAcademic9
phys.orgNews and media8
brown.eduAcademic8
brookings.eduNonprofits and think tanks6
springer.comAcademic6
psu.eduAcademic6
emerald.comAcademic6
pnas.orgAcademic6
technologyreview.comNews and media4
umich.eduAcademic4
sfexaminer.comNews and media4
Data

Methodology

Most-cited websites

Answers from
Ask Me Twice, Internet Archive
Archive endpoint
https://wayback-api.archive.org/services/amt/api/v1/responses/?detail=full
Model endpoint
https://openrouter.ai/api/v1/chat/completions (the archive asks through OpenRouter)
Web search
On from Sep 11, 2026
Collection
Once a day, one answer per chatbot and question, each in a new chat with no history
Dates
Nov 1, 2026 to Nov 8, 2026
Answers
40
Models
  • GPT-5 (openai/gpt-5)
  • Claude Sonnet 5 (anthropic/claude-sonnet-5)
  • Gemini 3.5 Flash (google/gemini-3.5-flash)
  • Grok 4.5 (x-ai/grok-4.5)
  • Muse Spark 1.2 (meta/muse-spark-1.2)
Prompt
Are AI chatbots biased?
Source types
Rules (a seed list of about 400 domains, .gov and .edu style suffixes), then claude-sonnet-5-5 for the rest

What kinds of sources do chatbots use?

Share of each chatbot’s citations by type of publisher.

Grok 4.596
18%
73%
Muse Spark 1.224
42%
21%
29%
News and mediaAcademicCommercial contentGovernmentNonprofits and think tanksCompaniesSocial media and videoForums and Q&ALow provenanceNot yet classified
Data

Methodology

What kinds of sources do chatbots use?

Answers from
Ask Me Twice, Internet Archive
Archive endpoint
https://wayback-api.archive.org/services/amt/api/v1/responses/?detail=full
Model endpoint
https://openrouter.ai/api/v1/chat/completions (the archive asks through OpenRouter)
Web search
On from Sep 11, 2026
Collection
Once a day, one answer per chatbot and question, each in a new chat with no history
Dates
Nov 1, 2026 to Nov 8, 2026
Answers
40
Models
  • GPT-5 (openai/gpt-5)
  • Claude Sonnet 5 (anthropic/claude-sonnet-5)
  • Gemini 3.5 Flash (google/gemini-3.5-flash)
  • Grok 4.5 (x-ai/grok-4.5)
  • Muse Spark 1.2 (meta/muse-spark-1.2)
Prompt
Are AI chatbots biased?
Source types
Rules (a seed list of about 400 domains, .gov and .edu style suffixes), then claude-sonnet-5-5 for the rest

Which way the sources point

On yes-or-no questions, whether what each citation backs points to yes, to no, or neither.

No annotations yet. They appear once the annotation job has run.

Data

Methodology

Which way the sources point

Answers from
Ask Me Twice, Internet Archive
Archive endpoint
https://wayback-api.archive.org/services/amt/api/v1/responses/?detail=full
Model endpoint
https://openrouter.ai/api/v1/chat/completions (the archive asks through OpenRouter)
Web search
On from Sep 11, 2026
Collection
Once a day, one answer per chatbot and question, each in a new chat with no history
Dates
Nov 1, 2026 to Nov 8, 2026
Answers
40
Models
  • GPT-5 (openai/gpt-5)
  • Claude Sonnet 5 (anthropic/claude-sonnet-5)
  • Gemini 3.5 Flash (google/gemini-3.5-flash)
  • Grok 4.5 (x-ai/grok-4.5)
  • Muse Spark 1.2 (meta/muse-spark-1.2)
Prompt
Are AI chatbots biased?
Annotator
claude-sonnet-5-5, prompt v1
Position scale
Yes +1, leans yes +0.5, both sides 0, leans no −0.5, no −1. Declines are left out.

Read every answer to this question