Reflection AI
- Type
- AI company
- Place
- United States — USA
- Last checked
- 2026-09-22
- Next check due
- 2027-03-21
This record is preserved as reference material. It is not a current contact destination and is excluded from Places to contact and recommendations.
Ways to reach them
- Information only · 2026-09-22
Single catch-all contact
- Recipient
- Reflection AI
- Who may use it
- People following the published contact instructions
- What it accepts
- Observed model behaviour; AI model flaws
Information only. This link provides background or evidence; it is not a place to send a message.No contact action is available for this entry. - Information only
Support knowledge base
- Recipient
- Reflection AI
- Who may use it
- People following the published contact instructions
- What it accepts
- Observed model behaviour; AI model flaws
Information only. This link provides background or evidence; it is not a place to send a message.No contact action is available for this entry. - Information only
Homepage
- Recipient
- Reflection AI
- Who may use it
- People following the published contact instructions
- What it accepts
- Observed model behaviour; AI model flaws
Information only. This link provides background or evidence; it is not a place to send a message.No contact action is available for this entry. - Information only
No published safety framework found
- Recipient
- Reflection AI
- Who may use it
- People following the published contact instructions
- What it accepts
- Observed model behaviour; AI model flaws
Information only. This link provides background or evidence; it is not a place to send a message.No contact action is available for this entry.
What it does
No publicly named safety or alignment function and no published safety framework, despite positioning as an open frontier lab.
Honest assessment
No public accounts of safety contact. Reflection raised at frontier-lab scale and positions itself as America's open-weights answer to Chinese labs, but publishes no safety framework, no disclosure policy and no security contact — only a privacy policy with one catch-all address. The fact that 'correction requests regarding AI model outputs' are handled as a privacy-law obligation rather than a safety process is the tell.
Notes
Record as: no meaningful public safety channel. Read the address off the privacy policy yourself; do not trust a transcribed version. Expect it to be handled as a data-subject request.