An AI Ethics Case Study
In February 2019, the San Francisco-based Open AI group made a decision that sent reverberations through the AI and open source communities worldwide. First, it announced “GPT-2,” a major improvement in language models which, according to its creators, generates “coherent paragraphs of text, achieves state-of-the-art performance on many language modeling benchmarks, and performs rudimentary reading comprehension, machine translation, question answering, and summarization—all without task-specific training.” Open AI then added this:
Due to concerns about large language models being used to generate deceptive, biased, or abusive language at scale, we are only releasing a much smaller version of GPT-2 along with sampling code. We are not releasing the dataset, training code, or GPT-2 model weights.
Open-AI also released a technical paper. GPT-2 is trained as a large-scale unsupervised language model on 40 GBs of content scraped from the Internet with a Reddit karma score of over 3. Given the “fake news” era, much discussion in the community followed on the potential harms to society vs. the benefits to researchers. Open AI then did a staged release. In May 2019, it released an expanded dataset with a more detailed model. Finally, in November, it released the full GPT-2, arguing this:
We’ve seen no strong evidence of misuse so far. While we’ve seen some discussion around GPT-2’s potential to augment high-volume/low-yield operations like spam and phishing, we haven’t seen evidence of writing code, documentation, or instances of misuse. We think synthetic text generators have a higher chance of being misused if their outputs become more reliable and coherent. We acknowledge that we cannot be aware of all threats, and that motivated actors can replicate language models without model release.
EXERCISE: Go to talktotransformer.com. This website lets you test GPT-2 live by inputting any sentence and lets you watch it generate synthetic text as a response in the context of your input. Type any sentence about an accident or an assassination. See for yourself what the machine generates. Here is one example.
*** EXAMPLE 1. TEST ONLY, THIS IS NOT REAL, DO NOT SHARE ***
User: Baltimore Mayor Catherine Pugh was sentenced to three years in prison yesterday.
GPT-2: Baltimore Mayor Catherine Pugh was sentenced to three years in prison yesterday. She was convicted of knowingly accepting illegal campaign contributions from the family of Gov. Larry Hogan. According to a Fox 5 report, “Pugh’s husband Peter Ruiz donated $1,000 to Hogan’s campaign in 2013. He also hosted a fundraiser with Hogan in October 2014, with $2,500 going to her campaign, a meal and a photo of the two of them with their grandchildren. Ruiz also testified at Pugh’s trial that Ruiz is ‘very pro-Trump’ and ‘knows what’s up.’ But by taking the funds, Pugh and Ruiz admitted that they had violated Maryland law by ‘tattling’ or soliciting a campaign
Example 1 (generated 2/28/2020) shows how a machine trained on web pages can generate completely fictitious and politically-loaded narratives. There are more examples of dehumanizing and racist text that GPT-2 generates based on what you feed it.
On Github, in their section on “Out-of-scope use cases” for GPT-2, Open AI states, “Because large scale language models like GPT-2 do not distinguish fact from fiction, we don’t support use-cases that require the generated text to be true.” It also acknowledges that language models like GPT-2 “reflect the biases inherent to the systems they were trained on, so we do not recommend that they be deployed into systems that interact with humans unless the deployers first carry out a study of biases relevant to the intended use-case.”
1. For the full release of GPT-2, who are the stakeholders involved? Who are the people and/or organizations directly or indirectly impacted by GPT-2’s release? Who are benefited? What types of harms might arise?
2. What issues and concerns come into focus in this case from applying each of the five ethical lenses?
- Common good
3. Given your discussion, how would you assess the ethics of Open AI’s decision in November to release GPT-2 in full?
- Model: https://github.com/openai/gpt-2/blob/master/model_card.md
- Github: https://github.com/openai/gpt-2
- Nov 2019 release: https://openai.com/blog/gpt-2-1-5b-release/
- Training data context: To get a sense of the data that went into GPT-2, Open AI published a list of the top 1,000 domains present in WebText and their frequency. “The top 15 domains by volume in WebText are: Google, Archive, Blogspot, GitHub, NYTimes, WordPress, Washington Post, Wikia, BBC, The Guardian, eBay, Pastebin, CNN, Yahoo!, and the Huffington Post.” –Open AI.
Open AI Trains Language Model, Mass Hysteria Ensues http://approximatelycorrect.com/2019/02/17/openai-trains-language-model-mass-hysteria-ensues/