6 New Studies Put AI to the Test
MIT scholars offer data-based evidence to guide employers, developers and policymakers
By Peter Krass
GenAI search results are distrusted. Job descriptions generated by large language models are efficient, but don’t always lead to more hires. And adding “beneficial friction” can help AI users to be more thoughtful and accurate.
These are some of the latest findings of research projects under way at the MIT Initiative on the Digital Economy (IDE). At the 2024 Annual Conference, an online event presented for members during the week of May 20, many speakers described ongoing AI studies they hope will provide practical guidance to those building and using GenAI.
The conference featured presentations by eight research group leaders, along with postdocs and doctoral students covering non-AI topics, as well.
Among the AI-related studies, topics included AI trust, marketing, policy and economics. All aimed at improving the quality and experiences with GenAI. Some highlights from these presentations include the following:
1. A Matter of AI Trust
Sinan Aral, director of the IDE and lead of its AI and Decentralization Group, described a new experiment involving nearly 5,000 users aimed at discovering whether people trust GenAI and, if so, under what conditions.
Use of GenAI search responses by Google and others is on the rise, but Aral and his fellow experimenters found that people have less trust in GenAI than with other information sources.
Trust in AI search can be a slippery slope. In the experiment, trust improved when GenAI demonstrated reliability, but decreased when the results were uncertain. Further, Aral explained, people sometimes trust AI when they shouldn’t. (See related blog, Sending GenAI into the Wild.)
2. Gaining ‘Beneficial’ Friction
AI was also on the agenda of Renée Richardson Gosline, head of the IDE’s Human First AI Group. Gosline discussed “beneficial friction,” which involves adding digital “speed bumps” to AI systems that encourage users to be more deliberative.
When people take more time to consider AI results, they’re also more likely to change and correct their misperceptions.
“Our goal is to amplify the benefits of AI,” Gosline told attendees, “and to minimize any potential harm.” (See related interview for more details.)
MIT postdoc Jerry Zhang explained findings from an experiment he and Gosline conducted to examine human perceptions of AI in advertising. When people don’t know whether content has been generated by humans or AI, they generally consider the AI-generated content to be more valuable, he said. But when they do know how the content was created, they favor the content created by humans. (For more, see the IDE Research Brief: How Do People Regard AI-Generated Content?)
3. A Job for AI?
John Horton, an Associate Professor at MIT Sloan and lead of the IDE’s AI Marketplaces and Labor Economics Group, discussed work on AI as a writing aid. The study was conducted with Emma Wiles, an MIT Sloan doctoral candidate.
In one experiment, GenAI wrote the first drafts of job descriptions that employers could post online to attract candidates. Employers with access to the AI-written drafts were about 20% more likely to post the descriptions than those without access, and they spent about 40% less time writing or editing job posts than did the control group.
However, when it came to hiring, the advantages were less apparent. Employers with access to AI-written job descriptions made nearly 20% fewer hires than others — an unexpected result. Access to AI also sidelined efforts that employers might have otherwise put into writing more specific job posts since many of the AI-generated job posts were generic, according to Wiles.
The lower hiring rate among the treatment group surprised the researchers who thought that using AI would both improve the descriptions and increase the number of hires.
Experimental results don’t always work out as planned, Horton said. “But it gives us a road map for how to improve these kinds of features, which we still think have an enormous amount of potential.” (Read the full research paper, More, But Worse.)
4. Data Provenance, Please
A relatively new AI topic known as data provenance was the subject of an Annual Conference presentation by Shayne Longpre, a doctoral candidate at the MIT Media Lab. He’s also among the founders of the Data Provenance Initiative, an effort to audit the datasets (1,800 so far) used to train LLMs. (See their position paper, The Data Provenance Institute: A Large Scale Audit of Dataset Licensing and Attribution in AI.)
The history of a dataset’s ownership might seem obscure, but it’s vital to the accuracy of AI training data. A dataset may be inappropriate for a given application or it may not represent the right tasks, topics, domains or languages. It may even be used illegally. “We realized this information was not well understood or documented,” Longpre said.
Some of the group’s audit results were disturbing. For one, they found that on HuggingFace, a major platform that hosts datasets, nearly two-thirds of the datasets (65%) had incorrect or omitted licenses that state access permissions.
To help other developers and users, the 10 founders of the initiative, including MIT, are developing the Data Provenance Explorer. This tool will let users select subsets of languages and licensing constraints, submit their selection across different criteria, and see information and statistics about the underlying data. (Read the related paper, Data Authenticity, Consent and Provenance for AI are All Broken.)
5. Regulating the Regulations
Another proactive AI project was described by Robert Mahari, a researcher pursuing a joint law-and-business doctorate at the MIT Media Lab and Harvard Law School; he also worked on the Data Provenance Initiative.
Mahari’s regulation by design concept involves embedding regulatory objectives directly into a technical design.
Mahari said that this type of deliberate design would give people the confidence to use AI systems without worries about potentially violating a law or regulation.
Mahari and his colleagues are working with nonacademic bodies to foster real-world implementations. These partners include the European Union’s AI Act commission, World Bank, U.S. Copyright Office and Singaporean Privacy Agency.
“Compliance and regulation by design represent a risk-management paradigm that’s uniquely suited for AI,” Mahari said. “Intelligent technology design can proactively prevent failures and risks.”
6. The Economic Advantage
Researchers from the IDE’s AI, Quantum and Beyond research group addressed the economic and policy repercussions accompanying AI’s rapid technological advancements.
Group leader Neil Thompson presented the findings of his recent paper; Which Tasks are Cost-Effective to Automate with Computer Vision? The study used AI computer vision to measure actual “job replacement we should expect” and to explain why some reports about AI-driven job loss are “a little overblown,” Thompson told attendees.
Importantly, businesses need to determine which tasks now done by humans are economically attractive to automate with AI.
While Thompson believes the interest and excitement around long-term AI deployments are warranted, he also expects “a much more gradual [adoption] as it takes longer for costs to go down and for deployments to scale.” (For more, see Thompson’s interview: New Research May Calm Some of the AI Job-Loss Clamor.)
IDE postdoc Nur Ahmed explained the results of another recently published paper, The Growing Influence of Industry in AI Research. Ahmed and his co-authors assert that AI research is being dominated by industry over academia in a way that should leave policymakers worried.
Building on Ahmed’s conclusions, Ana Trisovic, a research scientist at MIT’s Computer Science and AI Laboratory (CSAIL) and FutureTech, spoke about the democratization of AI. She questioned whether AI is currently accessible and usable by a broad enough range of people and organizations. This issue, she noted, has important implications for regulatory policy, research practices and societal equity.
Trisovic, basing her comments on large-scale experimental research, believes
“there is unequal access to computational resources and technologies that influences who can participate in this AI-driven economy.”
Such limited access, she added, “restricts scientific benefits and innovation potential to only a very few well-resourced institutions, creating a big disparity in research advancement.”
Do more:
- Learn more about the IDE
- Browse IDE research and publications
- Explore the IDE’s eight Research Groups
Peter Krass is a contributing writer and editor to the MIT IDE.
