This week an #AI model was released on @huggingface that produces harmful + discriminatory text and has already posted over 30k vile comments online (says it's author).
This experiment would never pass a human research #ethics board. Here are my recommendations.
@huggingface as the model custodian (an interesting new concept) should implement an #ethics review process to determine the harm hosted models may cause, and gate harmful models behind approval/usage agreements.
Medical research has functional models ie for data sharing.
2/7
Open science and software are wonderful principles but must be balanced against potential harm. Medical research has a strong ethics culture bc we have an awful history of causing harm to people, usually from disempowered groups.
Finally I'd like to talk about @ykilcher's experiment here. He performed human experiments without informing users, without consent or oversight. This breaches every principle of human research ethics.
4/7
Imagine the ethics submission!
Plan: to see what happens, an AI bot will produce 30k discriminatory comments on a publicly accessible forum with many underage users and members of the groups targeted in the comments. We will not inform participants or obtain consent.
5/7
AI research has just as much capacity to cause harm as medical research, but unfortunately even small attempts to manage these risks (such as @NeurIPSConf#ethics code of conduct: openreview.net/forum?id=zVoy8…) are bitterly resisted by even the biggest names in AI research.
6/7
Twitter ate my original thread so I'm one tweet short and don't know what I missed. Sorry.
7/7
UPDATE: @huggingface has removed the model from public access and will implement a gating feature.
Furthermore they are doing for community feedback on an appropriate ethics review mechanism. This is really important so please engage with this.
@huggingface should implement gated access to prevent harmful misuse of this model and others like it. Open science and software are great principles, but must be balanced against the risk of real harm. Medical research has functional models for this ie for data sharing.
2/7
Furthermore, @huggingface (as the model custodian, which is an interesting new concept) should be responsible to assess which models pose a risk. An internal review board process should be set up, with model authors required to submit a summary of risk for their models.
3/7
Additionally, I want to discuss @ykilcher's choice to perform an experiment on human participants without #ethics board approval, consent, or even their knowledge. This breaches every principle of human research ethics.
This has implications for AI research as a field.
4/7
Imagine the ethics proposal!
Plan: an AI bot which will post 30k+ discriminatory comments on a publically accessible forum often populated by underage users, including members of the marginalised groups the comments target. We will not inform them or ask for consent.
5/7
Medical research has strong ethics framework for a reason. We have a disgusting history of causing harm, particularly to disempowered people. See en.wikipedia.org/wiki/Medical_e… for examples.
AI has an equal capacity for harm, and this experiment was not uniquely harmful.
6/7
AI research desperately needs a ethical code of conduct, but even small steps like those taken by @NeurIPSConf (openreview.net/forum?id=zVoy8…) are bitterly fought against, by some of the most prominent researchers in the field.
AI researchers, do better. Please.
7/7
• • •
Missing some Tweet in this thread? You can try to
force a refresh
Very excited to have 2 new papers in press today in Lancet Digital Health, alongside an editorial from the journal highlighting our work.
I am immensely proud of the work we have done here and honestly think this is the most important work I have been involved in to date 🥳
1/7
#Medical#AI has a problem. Preclinical testing, including regulatory testing, does not accurately predict the risks that AI models pose once they are deployed in clinics.
I've written about this before in my blog, for example in:
We've put out a preprint reporting concerning findings. AI can do something humans can't: recognise the self-reported race of patients on x-rays. This gives AI a path to produce health disparities.
This is a big deal, so we wanted to do it right. We did dozens of experiments, replication at multiple labs, on numerous datasets and tasks.
We are releasing all the code, as well as new labels to identify racial identity for multiple public datasets.
2/8
Humans can't detect race better than chance, but AI performs absurdly well on the task. As you can see here, AUC scores are in the high 90s, and are maintained on external validation on completely distinct datasets and across multiple different imaging tasks.
Docs are ROCs: A simple fix for a methodologically indefensible practice in medical AI studies.
Widely used methods to compare doctors to #AI models systematically underestimate doctors, making the AI look better than it is! We propose a solution.
The most common method to estimate average human performance in #medical AI is to average sensitivity and specificity as if they are independent. They aren't though - they are inversely correlated on a curve.
The average points will *always* be inside the curve.
2/7
The only solution currently is to force doctors to rate images using confidence scores. While this works well in the few tasks where these scales are used in clinical practice, what does it mean to say you are 6/10 confident that there is a lung nodule?
Alright, let's do this once last time. Predictions vs probabilities. What should we give doctors when we use #AI / #ML models for decision making or decision support?
This discussion was getting long, so I thought I'd lay out my thoughts on a common argument: should models produce probabilities or decisions? Ie 32% chance of cancer vs "do a biopsy".
I favour the latter, because IMO it is both more useful and... more honest. IMO:
I personally suspect the biggest problem is automation bias, which is where the human over-relies on the model output.
Similar to self driving cars where jumping to complete automation appears to be safer than partial automation.
But interestingly (and perhaps counter-intuitively) this could also mean that "blind" ensembling (where the human gets no AI input, and the human and AI opinions are combined algorithmically) might be better than showing the doctor what the AI thinks.