Artificial intelligence (“AI”) is improving, but even the best models still can hallucinate, miscite, and miscalculate. The primary strategy for managing these and other risks associated with AI deployment is human review, also known as putting a “human-in-the-loop.” Here are various measures that we have seen businesses use to optimize human review of AI decisions and AI-generated content and ways that human review can effectively reduce risk without undermining the efficiencies gained from the AI use.
Human Expertise: Make sure that the human reviewer is an expert in the subject that they are overseeing. Humans can only be effective quality controls if they have sufficient knowledge, experience, expertise, and context to tell when something the AI generated is wrong, missing something important, or not fit-for-purpose.
Machine Review of AI: There are errors that even expert humans cannot easily detect without redoing most of the work. For example, a human may not be able to check a specific number that was generated by a complex calculation. To address this risk, some businesses are having one AI model review the work of another AI model and are also designing a workflow where certain AI outputs must be within a predetermined range of acceptable options, with the reviewing AI flagging anything beyond that range as likely to be in error and in need of human review.
Show Its Work: To facilitate review of AI output, where feasible, the AI should be configured to display the relevant portion of the source material in its output, rather than just providing a link to the source.
Review the Reviews: Implement compliance tools to help ensure that any required human review is both actually happening and effective. For example, companies can require that the human reviewer take active steps to affirmatively acknowledge that relevant citations have been checked (e.g., implementing checkboxes for each cite that, only once complete, enable the user to copy/paste the output). Companies can also record the length of time between checks to facilitate compliance reviews (e.g., automated flags are raised when a user completes checkboxes at a rate that is not consistent with meaningful human review of the referenced data). Some businesses are also routinely taking a small sample of human-approved AI decisions and having them reviewed again by an independent reviewer to assess error rates.
Human-Over-the-Loop: Some AI outputs do not require human review for each associated decision or use, which is usually what is meant by “human-in-the-loop.” Often, it is sufficient to allow the AI to make decisions on its own, but provide for human monitoring and spot checking of those decisions to ensure that the AI is behaving as expected, which is sometimes referred to as “human-over-the-loop.” An example of this kind of decision is whether a credit card purchase was fraudulent and the credit card should be disabled to avoid further fraud. This is usually a decision made by AI, which a human can quickly override if they can confirm that the AI’s decision was a false positive.
Human-Out-of-the-Loop: There are some circumstances where the AI should prevail over human decisions, which is referred to as “human-out-of-the-loop.” This usually arises in situations where there is a need for quick decisions to prevent significant harm and where machine decision-making is viewed as superior or there is a strong possibility that human decision-making is impaired. Examples include an AI-based cybersecurity detection tool that prevents a human from emailing out a malicious attachment that contains malware or factory machinery that shuts down if the AI monitoring system detects that its human operator is falling asleep or otherwise impaired.
Revisit Human Review Procedures Periodically: Procedures for human review of AI content and decision-making are new, and the optimal workflow can change quickly based on the improvements to the AI model, innovations in model supervision techniques, and changes in regulatory expectations. Periodic review of those procedures will help ensure that they remain efficient and effective in response to changing circumstances.
****
To subscribe to the Data Blog, please click here.
The cover art used in this blog post was generated by ChatGPT.
