Companies are reluctant to adopt AI solutions due to the difficulty of balancing the cost of governance with the behavior of large language models (LLM), such as hallucinations, data privacy violations, and the potential for the models to produce malicious content.
One of the most difficult challenges facing LLM adoption is specifying the model which is a damaging response, but IBM believes this can help improve the situation for companies around the world.
Elizabeth Daly, STSM, Research Manager, Interactive AI Group of IBM Research Europe, emphasized at an event in Zurich that the company wants to develop AI that developers can trust. She noted: “It's easy to measure and quantify clicks, but it's so easy to measure and quantify what constitutes malicious content.”
Detect, control, audit
Generic governance policies are not sufficient to control LLMs. Therefore, IBM aims to develop LLMS to use the law, corporate norms and the internal governance of each individual company as a control mechanism – allowing governance to go beyond corporate norms and integrate individual ethics and norms. social norms of the country, region or sector in which it is used.
These documents can provide context to an LLM and can be used to 'reward' an LLM for remaining relevant to their current job. This allows for an innovative level of fine-tuning in determining when AI produces harmful content that could violate a region's social norms, and could even allow an AI to detect whether its own output can be identified as harmful.
Additionally, IBM has been meticulous in developing its LLMs based on data that is reliable, detecting, monitoring, and checking for potential biases at every level, and has implemented detection mechanisms at every stage of the pipeline. This is in stark contrast to off-the-shelf basic models that are typically trained on biased data and even if this data is later removed, the biases can still resurface.
The proposed EU AI law will connect the governance of AI to the intentions of its users, and IBM argues that usage is a fundamental part of how it will govern its model, as some users can use the AI for summary tasks, and others for classification tasks. Daly argues that usage is therefore a “first-class citizen” in IBM's governance model.