In the legal profession, as in nearly every other aspect of life, it is difficult to escape discussions or references to Generative Artificial Intelligence (GAI). The discourse, unsurprisingly, has taken on science-fiction overtones. Adoption advocates compare GAI’s potential to the post-scarcity utopia of Ian M. Banks’s Culture, while skeptics see the worlds of Philip K. Dick abounding in uncertainty and paranoia, where technology frequently breaks down or has unintended consequences. Luckily for legal professionals, bankruptcy courts provide a useful arena to stress-test these competing visions. AI companies will either deliver genuine value capable of restructuring or reveal themselves to be Dick’s Biltong, replications that increasingly degrade.
As AI startups increasingly seek bankruptcy protection, legal commentators have addressed several recurring challenges AI companies bring with them to bankruptcy: AI washing, uncertain ownership under section 541, and restrictions on assumption and assignment under section 365. What has received less attention is a problem that distinguishes AI assets from conventional software: They decay. Although GAI dominates the conversation, the problem described herein affects AI models of all kinds, generative or otherwise. This article calls that phenomenon the Ubik problem and explains its implications for bankruptcy and how counsel can prepare for it.
In Philip K. Dick’s Ubik (1969), characters watch the world regress: Objects revert to earlier, obsolete forms unless sustained by the aerosol spray called Ubik. An AI model faces a similar problem. In AI models, this decay can take two forms: (1) Model Drift and (2) Model Collapse. Decay in Model Drift occurs because the world around the model has changed. Every AI model is a predictive machine that utilizes large swaths of data to predict the most probable outputs. As the environment changes, the predictive model is less representative of the environment although its underlying data has not changed. This can occur for a variety of reasons. For example, many card issuers’ fraud-detection models were trained on years of pre-2020 transactions, when most customers shopped in person and an online purchase from an unfamiliar merchant was a meaningful warning sign. When the Covid-19 pandemic moved most everyday spending online overnight, these models began flagging large volumes of legitimate purchases as suspicious. Until engineers retrained the models on current transactions, their predictions grew steadily less accurate. In the GAI context, a customer service chatbot is trained on data from a specific time. As products or policies change, the model continues to give confident answers that reflect old products and policies.
The other form of decay is Model Collapse. This occurs when an AI model slowly incorporates more synthetic, AI-generated data into its dataset. The synthetic data will contain errors from the AI model, which it will then learn from to produce more content. Eventually, the model will learn from nothing but errors and completely lose its usefulness.
Thus, to combat the Ubik problem, counsel must act with urgency on multiple fronts. First, AI models can be maintained, but they need engineers to maintain them. Prior to filing or early in the case, counsel should negotiate a Key Employee Retention Program (KERP) or other retention program to ensure that key engineers do not leave the company when the case is filed.
Second, to maximize the value of the AI model, counsel will need to navigate the myriad agreements that cover the different parts necessary to make the model work. Every AI model contains datasets, software, and infrastructure to deploy the model and contractual rights to access and use the underlying data. Counsel must navigate these byzantine arrangements quickly. This may require moving to a 365 process sooner rather than later or simply negotiating consent from different licensors before filing or early in the case. Note, getting consent may be necessary regardless, depending on which circuit you file in, as some circuits do not allow the debtor to assume the license if a hypothetical person could refuse to consent to the assignment of the contract under 365(n).
Third, secured creditors whose collateral includes AI assets may assert that model drift constitutes diminution in value and demand enhanced adequate protection. Counsel will need to anticipate this demand and address it early in the bankruptcy process. This may include proposing periodic model performance audits, offering additional collateral, or negotiating cash reserves to fund ongoing maintenance.
Fourth, prospective buyers will likely request technical diligence on model performance trends. Counsel should prepare for these requests on the front end by creating a data room for potential buyers that contains records showing the preservation of non-AI data sources, tracking the data lineage, and monitoring and maintenance of the model’s performance over time. These measures will allow for a formal process with proper diligence while compressing a 363-sale timeline.
Finally, speed is essential not only to combat the Ubik problem but also because the AI landscape evolves rapidly. A competitor’s model may surpass the debtor’s technology while the case proceeds, eroding relative market position and diminishing buyer interest. Working proactively on all fronts is therefore key to achieving the best outcome for the client.
Whether GAI delivers the post-scarcity promise of Banks’s Culture or the paranoid instability of Philip K. Dick remains to be seen. However, counsel who treat AI assets like conventional property will find themselves in Dick’s world, where the ground shifts beneath their feet and yesterday’s certainties no longer hold. Practitioners must become fluent in a new dialect, one where technical diligence, licensing analysis and speed are as important as statutory interpretation.
For more information or questions, contact Schuyler Pals.
Associate