AI Refusals and Honest Workarounds

Every few weeks someone in the group posts that an AI refused a perfectly ordinary request: translate a sicha, transcribe their own PDF, put a posuk to music. Most of these refusals are a filter matching a pattern (copyright, a real person's face, a safety topic) without understanding the situation. Most can be resolved honestly, by giving the context the filter was missing or by using a tool built for the job.

This page covers only legitimate work and legitimate fixes. It does not cover tricks for getting a model to do something its makers deliberately ruled out. Those break the terms of service, can get an account closed, and produce less reliable output. And sometimes the refusal is right: if the request is to dig up a private person's details, or to make a document look like an official original, stop.

Copyright refusals on Torah texts

What happened.

• November 2025: free ChatGPT refused to translate an uploaded sicha PDF, citing copyright.
• August 2026: a scheduled Claude Cowork task would not pull a parsha summary from Chabad.org "due to copyright." Pointing it to the site's copyright policy did not help.
• October 2026: Claude would not transcribe a PDF a member had written himself until he confirmed he owned it. He noted other assistants did it without asking.
• September 2026: Daiah, the library of the Rebbe's Torah now at chabad.ai, blocks copying with a notice that reuse needs written permission. That one is not a misfire; it is the publisher's choice.

What worked.

• Say what it is and why. "This is a sicha I am learning to prepare a class for my community. Translate it for my own study." If the work is yours, say so. Do not claim rights you do not have.
• Give text, not a file. Upload the PDF to Google Drive and open it with Google Docs, or use Google Lens, then paste the text. A plain-text paste is less likely to trip the filter than a scanned book page, and it fixes the two-column problem at the same time. Clean text sources are on Hebrew & Yiddish.
• Start a fresh chat. Once a model has refused in a conversation, the refusal stays in its context and colors what follows. Anthropic's own help page says the same about its filters: the original request stays in the conversation and can trigger the block again.
• Use the publisher's own channels. For a weekly summary from a site, use its link, newsletter or feed in your message instead of having an agent copy the page. For anything you want to republish, ask the publisher. The debate over how open the Rebbe's Torah should be to AI is on Chassidus, Copyright and Access.

A posuk in a song (Suno)

In August 2026 Suno refused a posuk as "copyrighted lyrics." The group admin's guess was that a well-known song uses the same words. That fits the timing: on August 6, 2026 Suno began screening prompts and output with Musixmatch's Sentinel, which matches lyrics against a database of songs (report). Honest options: write your own verses around the posuk, use a recording of a real nigun, or report the false match to Suno. See Suno.

Photos of the Rebbe, and of real people

What happened. In August 2025 a shliach found that every time he put a real photo of the Rebbe into an image tool, it altered the face, prompt after prompt. Image generators redraw a picture rather than edit it, and faces are what they change most. In August 2026 members found the same with group photos: ChatGPT and Gemini changed faces when asked only to remove backpacks.

What worked. Do not run a real photo of the Rebbe, or of anyone, through an image generator. Have the AI make only the background, frame or layout, then place the original photo on top in Canva. For small cleanups, Canva's Magic Eraser did best in members' tests.

The open question. In October 2026 a member wrote that he is "not a fan of AI making pics of Rabeim... accurate or not"; another defended one image as no more than a copy of a photograph. The group did not resolve it. It is a question for your mashpia or rav, not for a prompt.

While you are checking images: a member warned in October 2026 that AI Torah graphics sometimes add a cross to the crown of a Sefer Torah. Look before you print.

Pictures and video of your community

AI image and video tools often get Chabad dress and settings wrong. In November 2025 a shliach's video prompt came back with clothing that looked nothing like his community. The fix the group settled on is to describe what should be visible (the clothing, the setting, the activity) instead of naming a group. That thread is on Depicting Your Community. The video tool from that thread, OpenAI's Sora app, was shut down: OpenAI announced its discontinuation on March 24, 2026.

Safety filters that misfire (Claude Fable)

What happened. In July 2026 a member saw Claude announce that Fable's safeguards had flagged his message, which it said "may flag safe and routine coding, cybersecurity, or biology work," and switch him to an older Opus model mid-chat.

What Anthropic says (help page): Fable 5 and Fable 5.1 run automated checks for offensive cybersecurity, dangerous biology and a few other areas. A blocked request is re-run on Opus (Opus 5 for biology and chemistry, Opus 4.8 for offensive cybersecurity), and the chat stays on Opus until you switch back. The checks also read your memory, connected apps, search results and uploaded files, so a pasted document can trigger a block you did not type.

What helps.

• Edit the message that was flagged and retry, instead of asking again below it.
• If you would rather have it stop than switch models, turn off "Switch models when a message is flagged" under Settings > Capabilities.
• If the block was clearly wrong, use "Send feedback" so Anthropic sees it.
• Most of the time, the Opus answer is fine for ordinary work. Read it and move on.

Refusals about people

• Giving capacity. In October 2025 Perplexity refused to research a person's net worth as "personal information." A member's approach was to ask the model which concern blocked it and then explain the context truthfully, for example that he was researching his own organization. For donor research, public records are the honest route: a foundation's IRS filings, for instance, are searchable on ProPublica's Nonprofit Explorer. Keep private individuals out of it. See Fundraising.
• Contact details. In January 2026 a shliach used Gemini to find the contact details of someone who had declined to give them. The group's response was that it was creepy and not something to do. When a person has said no, an AI that finds the information anyway is not a workaround.
• Summarizing the group. Meta AI inside WhatsApp refused to summarize the group chat (November 2025). Members export the chat instead (you only get messages from after you joined) and ask a chat tool about it. Remember the export holds other people's names and numbers; keep it private. See Privacy and Data.

Documents that look official

In July 2026 a shliach noted that Claude may refuse to clean up a scanned document so it "looks original," treating it as possible fraud. If you are restoring your own old flyer or program for a reprint, say so. If the document is a certificate, receipt, letter or anything someone else issued, the refusal is right.

A short checklist

When an AI refuses legitimate work:

1. Read the refusal. It usually names the rule (copyright, a real person, a safety topic).
2. Add the honest context it was missing: who you are, what the text is, what you will do with it.
3. Paste text instead of uploading the book page or giving a link.
4. Start a fresh chat, or edit the original message, rather than arguing under it.
5. Try the tool built for the job: a Torah library for sources, Canva for photos, a transcription service for audio.
6. If the honest answer is that you do not have permission, ask the person or publisher who does.

Related

• Chassidus, Copyright and Access
• Depicting Your Community
• Privacy and Data
• Claude
• Image Generation