TLDR: Focused Language Models (FLMs) are being championed by experts like FICO’s Scott Zoldi as a critical solution to the pervasive issue of hallucinations in Generative AI (GenAI). This approach offers greater control over training data, addressing concerns raised by major financial institutions and regulators about the risks associated with unchecked GenAI adoption.
The financial services sector is increasingly turning to Focused Language Models (FLMs) to combat the persistent challenge of ‘hallucinations’ in Generative Artificial Intelligence (GenAI). Scott Zoldi, a prominent voice in the field, highlights FLMs as a crucial mechanism to ensure the responsible operationalization of GenAI, a technology that has drawn significant scrutiny for its propensity to generate inaccurate or fabricated information.
Hallucinations in GenAI have been a focal point of concern, making headlines in courtrooms and across news platforms throughout the past year. Major Wall Street firms, including Goldman Sachs Group Inc., Citigroup Inc., and JPMorgan Chase & Co., have explicitly warned investors in their 2024 annual reports about new risks stemming from the escalating use of AI. These risks encompass not only software hallucinations but also potential employee morale issues, exploitation by cybercriminals, and the complexities of evolving global regulations.
Regulatory bodies are also voicing apprehension. Michael Barr, formerly the US Federal Reserve Bank’s vice chair for supervision, previously cautioned in February about the competitive pressures driving an aggressive adoption of GenAI within financial services. Barr articulated that such pressures could heighten governance, alignment, and financial risks associated with AI integration.
In response to these growing concerns, companies like FICO are advocating for a proactive approach. FICO champions the use of Focused Language Models (FLMs) and focused task models as a robust solution to thwart hallucinations before they manifest. This methodology fundamentally differs from commercially available Large Language Models (LLMs) and Small Language Models (SLMs), which typically offer limited control over the data used for their training.
FLMs, conversely, provide critical transparency and control over the quality and appropriateness of the data on which a core domain-focused language model is built. This foundational control allows users to subsequently create highly specific, task-oriented FLMs with tightly defined vocabularies and training contexts, thereby significantly reducing the likelihood of generating erroneous outputs.
Detecting hallucinations in general-purpose LLMs remains a formidable challenge due to their often uninterpretable algorithms, which rarely provide clear justifications for their responses. Even when Retrieval Augmented Generation (RAG) contexts are referenced, human inspection may reveal that the information was not genuinely utilized in the output. Experts suggest that the most effective strategy to minimize hallucinations is for organizations to build their own pre-trained fundamental generative AI models, often leveraging focused-domain and task-based approaches.
The real-world impact of GenAI hallucinations is stark. Research from Stanford University last year indicated that general-purpose GenAI tools like ChatGPT exhibited an error rate as high as 82% when applied to legal tasks. While purpose-built AI tools for legal applications performed better, they still produced hallucinations 17% of the time, necessitating rigorous and time-consuming scrutiny. A critical exacerbating factor across all industries is the human element: users may fail to detect or validate hallucinations, leading them to act directly upon incorrect AI-generated information.
Also Read:
- Generative AI Revolutionizes Indian Banking: Boosting Efficiency and Customer Experience
- Anthropic Unveils Specialized Claude AI for Financial Services, Featuring Enhanced Data Integration and Capabilities
By embracing focused language models, the industry aims to ensure that the ‘golden age of AI’ continues to shine brightly, mitigating the risks and fostering trust in this transformative technology.
