Platforms Dream of Electronic Shepherds? Combatting Online Harms Through Innovation (FTC)

Editor’s Note: From time to time, ComplexDiscovery highlights publicly available or privately purchasable announcements, content updates, and research from cyber, data, and legal discovery providers, research organizations, and ComplexDiscovery community members. While ComplexDiscovery regularly highlights this information, it does not assume any responsibility for content assertions.

To submit recommendations for consideration and inclusion in ComplexDiscovery’s cyber, data, and legal discovery-centric service, product, or research announcements, contact us today.

Background Note: Recently the Federal Trade Commission issued a report to Congress warning about using artificial intelligence (AI) to combat online problems and urging policymakers to exercise “great caution” about relying on it as a policy solution. The use of AI, particularly by big tech platforms and other companies, comes with limitations and problems of its own. The report outlines significant concerns that AI tools can be inaccurate, biased, and discriminatory by design and incentivize relying on increasingly invasive forms of commercial surveillance. The report may be beneficial for cybersecurity, information governance, and legal discovery professionals seeking to better understand the concerns and challenges of combatting online harms as outlined and accentuated by the Federal Trade Commission.

Federal Trade Commission Report*

Combatting Online Harms Through Innovation – FTC Report to Congress

Executive Summary Extract

The deployment of AI tools intended to detect or otherwise address harmful online content is accelerating. Largely within the confines — or via funding from — the few big technology companies that have the necessary resources and infrastructure, AI tools are being conceived, developed, and used for purposes including combat against many of the harms listed by Congress. Given the amount of online content at issue, this result appears to be inevitable, as a strictly human alternative is impossible or extremely costly at scale.

Nonetheless, it is crucial to understand that these tools remain largely rudimentary, have substantial limitations, and may never be appropriate in some cases as an alternative to human judgment. Their use — both now and in the future — raises a host of persistent legal and policy concerns. The key conclusion of this report is thus that governments, platforms, and others must exercise great caution in either mandating the use of, or over-relying on, these tools even for the important purpose of reducing harms. Although outside of our scope, this conclusion implies that, if AI is not the answer and if the scale makes meaningful human oversight infeasible, we must look at other ways, regulatory or otherwise, to address the spread of these harms.

A central failing of these tools is that the datasets supporting them are often not robust or accurate enough to avoid false positives or false negatives. Part of the problem is that automated systems are trained on previously identified data and then have problems identifying new phenomena (e.g., misinformation about COVID-19). Mistaken outcomes may also result from problems with how a given algorithm is designed. Another issue is that the tools use proxies that stand in for some actual type of content, even though that content is often too complex, dynamic, and subjective to capture, no matter what amount and quality of data one has collected. In fact, the way that researchers classify content in the training data generally includes removing complexity and context — the very things that in some cases the tools need to distinguish between content that is or is not harmful. These challenges mean that developers and operators of these tools are necessarily reactive and that the tools — assuming they work — need constant adjustment even when they are built to make their own adjustments.

The limitations of these tools go well beyond merely inaccurate results. In some instances, increased accuracy could itself lead to other harms, such as enabling increasingly invasive forms of surveillance. Even with good intentions, their use can also lead to exacerbating harms via bias, discrimination, and censorship. Again, these results may reflect problems with the training data (possibly chosen or classified based on flawed judgments or mislabeled by insufficiently trained workers), the algorithmic design, or preconceptions that data scientists introduce inadvertently. They can also result from the fact that some content is subject to different and shifting meanings, especially across different cultures and languages. These bad outcomes may also depend on who is using the tools and their incentives for doing so, and on whether the tool is being used for a purpose other than the specific one for which it was built.

Further, as these AI tools are developed and deployed, those with harmful agendas — whether adversarial nations, violent extremists, criminals, or other bad actors — seek actively to evade and manipulate them, often using their own sophisticated tools. This state of affairs, often referred to as an arms race or cat-and-mouse game, is a common aspect of many kinds of new technology, such as in the area of cybersecurity. This unfortunate feature will not be going away, and the main struggle here is to ensure that adversaries are not in the lead. This task includes considering possible evasions and manipulations at the tool development stage and being vigilant about them after deployment. However, this brittleness in the tools — the fact that they can fail with even small modifications to inputs — may be an inherent flaw.

While AI continues to advance in this area, including with existing government support, all of these significant concerns suggest that Congress, regulators, platforms, scientists, and others should exercise great care and focus attention on several related considerations.

Read the announcement of the report.

Read the Complete Report: Combatting Online Harms Through Innovation (PDF) – Mouseover to Scroll

Read the original report.

*Shared with permission.

Additional Reading

• Investments in eDiscovery (A Running Listing)
• An Abridged Look at the Business of eDiscovery: Mergers, Acquisitions, and Investments

Source: ComplexDiscovery