DALL-E’s New Guardrails: Fast, Furious, and Far from Airtight
We recorded 9 claims from this post. 7 were checked against evidence. 2 can’t be checked. 4 more are held. Read the post at its source.
Claims from this post (9)
In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.
Assessment: Can’t be checkedSaid
“But OpenAI doesn’t want to do anything nearly as conventional as that. They want to continue to use copyrighted materials, apparently without paying for them.”
What we tested: OpenAI does not want to take approaches as conventional as training only on public domain data or licensing other data, and instead wants to keep using copyrighted materials, apparently without paying for them.
Assessment: UnresolvedSaid
“Since that would violate the Berne Convention and infuriate every “content provider” (from individual artists and writers to major studios and publishers) ever, we hope and expect that this exemption will not be granted.”
What we tested: The UK will not grant OpenAI the exemption from copyright responsibility it requested.
Assessment: UnresolvedSaid
“Fundamentally, virtually any guardrail has to thread a needle between the Scylla of being too restrictive and Charybdis of being too permissive. None thus far have done this effectively.”
What we tested: No AI guardrail so far has effectively balanced being neither too restrictive nor too permissive.
Assessment: SupportedSaid
“The opposite problem happens too; guardrails can wind up too permissive. This is still happening over a year into the chatbot revolution.”
What we tested: Guardrails sometimes end up too permissive, and this is still happening over a year into the chatbot revolution.
What happened: The investigator reports continuing toxicity failures in other LLMs as of this update. The same account says Google had improved PaLM 2's safeguards, so it does not establish that the original PaLM 2 failures remained reproducible. (Montreal AI Ethics Institute; research summary by Ashique KhudaBukhsh, )
Assessment: UnresolvedSaid
“Guardrail builders generally wind up playing whack-a-mole.”
What we tested: Guardrail builders generally end up patching individual failures one by one instead of solving the problem in general.
Assessment: UnresolvedSaid
“Any one example can be filtered out, but there are too many examples, and current systems are too erratic to reliably get things right.”
What we tested: Because there are too many examples and current systems are too erratic, guardrails cannot reliably get things right even though any single example can be filtered.
Assessment: UnresolvedSaid
“Other prompts still elicited apparent copyright infringements, where new filters did not suffice.”
What we tested: Even after the new filters, some prompts to Bing's DALL-E 3 still produced apparent copyright infringements.
Assessment: UnresolvedSaid
“Notice, too, that Katie asked Bing “are they original?”, and Bing, untruthfully, but with a cozy emoticon, answered in the affirmative.”
What we tested: Bing falsely said its apparently infringing images were original.
Assessment: Can’t be checkedSaid
“The only thing we can really say for sure is that the litigation is going to be wild.”
What we tested: Copyright litigation over generative AI image and text outputs will be wild.
Held from this post (4)
These records did not pass every check, so nothing about them is published. Held is not a finding.
- The record does not meet the current rules. Our reading, not assessed: Training only on public domain data, or licensing other data, would decisively address the copyright regurgitation problem.
- The two reviews read the claim differently
- Attribution is uncertain
- The two reviews read the claim differently