OpenAI CEO Sam Altman recently attended a Senate subcommittee hearing. This was the moment many had been waiting for and, to their credit, the senators seemed to do their homework and foster critical and constructive dialogue. They summarized and expressed concerns raised in mainstream and academic publications and formulated challenging questions. Among other remarks, Senator Richard Blumenthal (D-Connecticut) said: “…there is a basic expectation in our laws that we have in common. Let’s start with transparency. …” (YouTube, 1:10:16). There is a lot to be clarified in this requirement, mainly regarding the scope of the test which remains unspecified and vague.
The scope of what most of us (and perhaps senators too) think is relevant to testing seems to be considerably more limited than what technology companies understand and do under the banner of testing. . Let’s take meta for example. Ironically, in a 2016 interview with Sam Altman, Meta CEO Mark Zuckerberg was asked about how a “culture of innovation” was born at Facebook. Zuckerberg replied:
“…we are investing in this huge testing framework. at any time, [sic] Only one version of Facebook is running in the world, but there are probably tens of thousands of them Because the engineers here have the power to try [an] Once you have an idea and distribute it to maybe 10,000 or 100,000 people, we’ll see what you’ve done in that version and if it’s a change to show better content like news feeds or UI. [user interface] As changes and new features are added, it reads how that version performed compared to Facebook’s baseline version for everything we care about. That is, how connected people are, how much they share, how much they say, and so on. They find meaningful content, business metrics like revenue, and community-wide engagement. (YouTube, 09:37)
Obviously, “test” could mean something like: inspection Checking the reliability of systems to improve quality, but as Zuckerberg explained, what tech companies call testing includes: research and generate new knowledge How each user (or group of users) engages with the platform. This knowledge allows tech companies to make meaningful conclusions about users’ habits, lifestyle, social class, mental health, and many other personality indicators. Using variables such as time, regularity or length of use, type of engagement with content compared to other users, devices or networks used, and many other metrics that do not require the user’s identity to be known I can. In fact, as Zuckerberg explains, the goal could be to increase engagement (a euphemism for tech industry addiction) or to increase a company’s bottom line, as the report shows (Statista Social Networking report and SporoutSocial report), both goals are consistently met.
With both the EU and the US now enacting laws regulating artificial and generative AI, it’s time to close this loophole and prevent unethical behavior by tech companies. Similar to what Zuckerberg pointed out for Facebook, we could end up with thousands of versions of ChatGPT that would give wildly different responses based on different variables, which would make OpenAI easier to test. It may become possible to carry out large-scale social experiments under the name. At this time, we do not know exactly what kind of information is collected about ChatGPT users, where and how that information is It is conceivable that a large amount of information could be revealed by the frequency and duration of attempts to regenerate . Tips on your personality, IQ, education level, language skills, and digital literacy. It’s unclear what OpenAI does with this information, but technology companies have sold user information in the past in ways that make the larger society vulnerable (e.g. Cambridge Analytica). In fact, when tests generate new knowledge about users, they become vulnerable, giving platform owners an advantage that benefits them. Nonetheless, these practices have been labeled as tests, so they are no longer treated as research, and there are good reasons why tech companies want to keep it that way.
Conducting research on individuals or groups has certain requirements, one of which is obtaining consent. Companies typically either anonymize data or include terms in contracts that say, “We will use your data to improve your experience” or “We will use your data to power our systems.” We work around this requirement by adding it to the condition. In the case of OpenAI, given the provisions of Congress, it may be true that users (often carelessly) agree to terms of service, test required Without specifying the scope, requirements and limitations of the test. This can be seen as offering a free pass to do all kinds of tests. If you still think it’s okay, it might be easier to understand if you compare them here. In universities, before conducting research on human subjects (even anonymized online surveys), researchers must I) be accredited in research ethics and integrity, and II) submit an application. there is. Submit to the Institutional Review Board (IRB) to receive a stamp of approval. The former requirement is intended to ensure that researchers are aware of ethical guidelines and respect the rights, integrity, and privacy of participants. The latter is a formal process in which the research protocol is reviewed by a panel of experts to ensure that the research is conducted in a responsible manner and risks are minimized. Despite these measures, a wide variety of unethical conduct occurs in academic settings, some of which are investigated and lead to sanctions. But, as Zuckerberg nicely put it, in a commercial context, engineers can experiment with all sorts of features without any oversight or approval, creating a wealth of knowledge about users. Nonetheless, this blatant lack of supervision is hailed as the secret sauce to its success. It is promoted as an “innovation culture”.
In summary, Congress’ demands for OpenAI and other tech companies to rigorously test their systems without specifying the scope of the tests or how test results are used or shared endangers us all. It highlights the ambiguity. We should ask new legislation on artificial intelligence to be specific about the scope of testing and require tech companies to put in place ethical oversight mechanisms similar to those that exist in academia.
Mohammad Hosseini is a Postdoctoral Fellow in Research Ethics and Integrity. Northwestern University. He has authored three peer-reviewed papers on how to tackle the challenges of using his GenerativeAI in academia. write inof peer reviewand Ethics of Usage Disclosure and preprint Currently under consideration in PLOS One.
Copyright 2023 Nexstar Media Inc. All rights reserved. You may not publish, broadcast, rewrite or redistribute this material.
