Login
Smart AI Policy Means Examining Its Real Harms and Benefits
The phrase "artificial intelligence" has been around for a long time, covering everything from computers with "brains"—think Data from Star Trek or Hal 9000 from 2001: A Space Odyssey—to the autocomplete function that too often has you sending emails to the wrong person. It's a term that sweeps a wide array of uses into it—some well-established, others still being developed.
Recent news shows us a rapidly expanding catalog of potential harms that may result from companies pushing AI into every new feature and aspect of public life—like the automation of bias that follows from relying on a backward-looking technology to make consequential decisions about people's housing, employment, education, and so on. Complicating matters, the computation needed for some AI services requires vast amounts of water and electricity, leading to sometimes difficult questions about whether the increased fossil fuel use or consumption of water is justified.
We are also inundated with advertisements and exhortations to use the latest AI-powered apps, and with hype insisting AI can solve any problem.
Obscured by this hype, there are some real examples of AI proving to be a helpful tool. For example, machine learning is especially useful for scientists looking at everything from the inner workings of our biology to cosmic bodies in outer space. AI tools can also improve accessibility for people with disabilities, facilitate police accountability initiatives, and more. There are reasons why these problems are amenable to machine learning and why excitement over these uses shouldn’t translate into a perception that just any language model or AI technology possesses expert knowledge or can solve whatever problem it’s marketed as solving.
EFF has long fought for sensible, balanced tech policies because we’ve seen how regulators can focus entirely on use cases they don’t like (such as the use of encryption to hide criminal behavior) and cause enormous collateral harm to other uses (such as using encryption to hide dissident resistance). Similarly, calls to completely preempt state regulation of AI would thwart important efforts to protect people from the real harms of AI technologies. Context matters. Large language models (LLMs) and the tools that rely on them are not magic wands—they are general-purpose technologies. And if we want to regulate those technologies in a way that doesn’t shut down beneficial innovations, we have to focus on the impact(s) of a given use or tool, by a given entity, in a specific context. Then, and only then, can we even hope to figure out what to do about it.
So let’s look at the real-world landscape.
AI’s Real and Potential Harms
Thinking ahead about potential negative uses of AI helps us spot risks. Too often, the corporations developing AI tools—as well as governments that use them—lose sight of the real risks, or don’t care. For example, companies and governments use AI to do all sorts of things that hurt people, from price collusion to mass surveillance. AI should never be part of a decision about whether a person will be arrested, deported, placed into foster care, or denied access to important government benefits like disability payments or medical care.
There is too much at stake, and governments have a duty to make responsible, fair, and explainable decisions, which AI can’t reliably do yet. Why? Because AI tools are designed to identify and reproduce patterns in data that they are “trained” on. If you train AI on records of biased government decisions, such as records of past arrests, it will “learn” to replicate those discriminatory decisions.
And simply having a human in the decision chain will not fix this foundational problem. Studies have shown that having a human “in the loop” doesn’t adequately correct for AI bias, both because the human tends to defer to the AI and because the AI can provide cover for a biased human to ratify decisions that agree with their biases and override the AI at other times.
These biases don’t just arise in obvious contexts, like when a government agency is making decisions about people. It can also arise in equally life-affecting contexts like medical care. Whenever AI is used for analysis in a context with systemic disparities and whenever the costs of an incorrect decision fall on someone other than those deciding whether to use the tool. For example, dermatology has historically underserved people of color because of a focus on white skin, with the resulting bias affecting AI tools trained on the existing and biased image data.
These kinds of errors are difficult to detect and correct because it’s hard or even impossible to understand how an AI tool arrives at individual decisions. These tools can sometimes find and apply patterns that a human being wouldn't even consider, such as basing diagnostic decisions on which hospital a scan was done at. Or determining that malignant tumors are the ones where there is a ruler next to them—something that a human would automatically exclude from their evaluation of an image. Unlike a human, AI does not know that the ruler is not part of the cancer.
Auditing and correcting for these kinds of mistakes is vital, but in some cases, might negate any sort of speed or efficiency arguments made in favor of the tool. We all understand that the more important a decision is, the more guardrails against disaster need to be in place. For many AI tools, those don't exist yet. Sometimes, the stakes will be too high to justify the use of AI. In general, the higher the stakes, the less this technology should be used.
We also need to acknowledge the risk of over-reliance on AI, at least as it is currently being released. We've seen shades of a similar problem before online (see: "Dr. Google"), but the speed and scale of AI use—and the increasing market incentive to shoe-horn “AI” into every business model—have compounded the issue.
Moreover, AI may reinforce a user’s pre-existing beliefs—even if they’re wrong or unhealthy. Many users may not understand how AI works, what it is programmed to do, and how to fact check it. Companies have chosen to release these tools widely without adequate information about how to use them properly and what their limitations are. Instead they market them as easy and reliable. Worse, some companies also resist transparency in the name of trade secrets and reducing liability, making it harder for anyone to evaluate AI-generated answers.
Other considerations may weigh against AI uses are its environmental impact and potential labor market effects. Delving into these is beyond the scope of this post, but it is an important factor in determining if AI is doing good somewhere and whether any benefits from AI are equitably distributed.
Research into the extent of AI harms and means of avoiding them is ongoing, but it should be part of the analysis.
AI’s Real and Potential Benefits
However harmful AI technologies can sometimes be, in the right hands and circumstances, they can do things that humans simply can’t. Machine learning technology has powered search tools for over a decade. It’s undoubtedly useful for machines to help human experts pore through vast bodies of literature and data to find starting points for research—things that no number of research assistants could do in a single year. If an actual expert is involved and has a strong incentive to reach valid conclusions, the weaknesses of AI are less significant at the early stage of generating research leads. Many of the following examples fall into this category.
Machine learning differs from traditional statistics in that the analysis doesn’t make assumptions about what factors are significant to the outcome. Rather, the machine learning process computes which patterns in the data have the most predictive power and then relies upon them, often using complex formulae that are unintelligible to humans. These aren’t discoveries of laws of nature—AI is bad at generalizing that way and coming up with explanations. Rather, they’re descriptions of what the AI has already seen in its data set.
To be clear, we don't endorse any products and recognize initial results are not proof of ultimate success. But these cases show us the difference between something AI can actually do versus what hype claims it can do.
Researchers are using AI to discover better alternatives to today’s lithium-ion batteries, which require large amounts of toxic, expensive, and highly combustible materials. Now, AI is rapidly advancing battery development: by allowing researchers to analyze millions of candidate materials and generate new ones. New battery technologies discovered with the help of AI have a long way to go before they can power our cars and computers, but this field has come further in the past few years than it had in a long time.
AI Advancements in Scientific and Medical ResearchAI tools can also help facilitate weather prediction. AI forecasting models are less computationally intensive and often more reliable than traditional tools based on simulating the physical thermodynamics of the atmosphere. Questions remain, though about how they will handle especially extreme events or systemic climate changes over time.
For example:
- The National Oceanic and Atmospheric Administration has developed new machine learning models to improve weather prediction, including a first-of-its-kind hybrid system that uses an AI model in concert with a traditional physics-based model to deliver more accurate forecasts than either model does on its own. to augment its traditional forecasts, with improvements in accuracy when the AI model is used in concert with the physics-based model.
- Several models were used to forecast a recent hurricane. Google DeepMind’s AI system performed the best, even beating official forecasts from the U.S. National Hurricane Center (which now uses DeepMind’s AI model).
Researchers are using AI to help develop new medical treatments:
- Deep learning tools, like the Nobel Prize-winning model AlphaFold, are helping researchers understand protein folding. Over 3 million researchers have used AlphaFold to analyze biological processes and design drugs that target disease-causing malfunctions in those processes.
- Researchers used machine learning simulate and computationally test a large range of new antibiotic candidates hoping they will help treat drug-resistant bacteria, a growing threat that kills millions of people each year.
- Researchers used AI to identify a new treatment for idiopathic pulmonary fibrosis, a progressive lung disease with few treatment options. The new treatment has successfully completed a Phase IIa clinical trial. Such drugs still need to be proven safe and effective in larger clinical trials and gain FDA approval before they can help patients, but this new treatment for pulmonary fibrosis could be the first to reach that milestone.
- Machine learning has been used for years to aid in vaccine development—including the development of the first COVID-19 vaccines––accelerating the process by rapidly identifying potential vaccine targets for researchers to focus on.
AI technologies can improve accessibility for people with disabilities. But, as with many uses of this technology, safeguards are essential. Many tools lack adequate privacy protections, aren’t designed for disabled users, and can even harbor bias against people with disabilities. Inclusive design, privacy, and anti-bias safeguards are crucial. But here are two very interesting examples:
- AI voice generators are giving people their voices back, after losing their ability to speak. For example, while serving in Congress, Rep. Jennifer Wexton developed a debilitating neurological condition that left her unable to speak. She used her cloned voice to deliver a speech from the floor of the House of Representatives advocating for disability rights.
- Those who are blind or low-vision, as well as those who are deaf or hard-of-hearing, have benefited from accessibility tools while also discussing their limitations and drawbacks. At present, AI tools often provide information in a more easily accessible format than traditional web search tools and many websites that are difficult to navigate for users that rely on a screen reader. Other tools can help blind and low vision users navigate and understand the world around them by providing descriptions of their surroundings. While these visual descriptions may not always be as good as the ones a human may provide, they can still be useful in situations when users can’t or don’t want to ask another human to describe something. For more on this, check out our recent podcast episode on “Building the Tactile Internet.”
When there is a lot of data to comb through, as with police accountability, AI is very useful for researchers and policymakers:
- The Human Rights Data Analysis Group used LLMs to analyze millions of pages of records regarding police misconduct. This is essentially the reverse of harmful use cases relating to surveillance; when the power to rapidly analyze large amounts of data is used by the public to scrutinize the state there is a potential to reveal abuses of power and, given the power imbalance, very little risk that undeserved consequences will befall those being studied.
- An EFF client, Project Recon, used an AI system to review massive volumes of transcripts of prison parole hearings to identify biased parole decisions. This innovative use of technology to identify systemic biases, including racial disparities, is the type of AI use we should support and encourage.
It is not a coincidence that the best examples of positive uses of AI come in places where experts, with access to infrastructure to help them use the technology and the requisite experience to evaluate the results, are involved. Moreover, academic researchers are already accustomed to explaining what they have done and being transparent about it—and it has been hard won knowledge that ethics are a vital step in work like this.
Nor is it a coincidence that other beneficial uses involve specific, discrete solutions to problems faced by those whose needs are often unmet by traditional channels or vendors. The ultimate outcome is beneficial, but it is moderated by human expertise and/or tailored to specific needs.
Context Matters
It can be very tempting—and easy—to make a blanket determination about something, especially when the stakes seem so high. But we urge everyone—users, policymakers, the companies themselves—to cut through the hype. In the meantime, EFF will continue to work against the harms caused by AI while also making sure that beneficial uses can advance.
