Sin embargo, a la hora de profundizar en las posibilidades y el potencial de la IA generativa, es fundamental examinar los grandes modelos lingüísticos (LLM) subyacentes que resultan tan influyentes en la continua evolución de esta tecnología. Los prejuicios, la desinformación y los estereotipos perjudiciales son sólo algunos de los ejemplos que pueden perpetuar, a menudo sin saberlo, los usuarios inconscientes. Por lo tanto, las organizaciones deben establecer directrices y políticas para garantizar que estos programas LLM se formen en conjuntos de datos éticos, y que la IA se despliegue de manera responsable para que no sean culpables de ningún discurso peligroso que pueda dañar su marca y reputación.
Los sesgos y alucinaciones perpetuados por la IA también tienen un impacto que va más allá de los contenidos perjudiciales y de dañar la reputación de las organizaciones. Las alucinaciones y los sesgos cognitivos transferidos a través de conjuntos de datos de entrenamiento también pueden crear vulnerabilidades de seguridad, afectando a las capacidades de una IA para reconocer con precisión las amenazas.
Biases and hallucinations perpetuated by AI also have an impact beyond harmful content and damaging organization reputations. Hallucinations and cognitive biases transferred through training data sets can also create security vulnerabilities, affecting an AI’s abilities to accurately recognize threats.
LLMs operate on transformer models within neural networks, making the tasks of narrative creation, language translation, and generating ‘human-esque’ content almost instinctive. LLMs can also analyze large volumes of unstructured data across an organization’s infrastructure, from numerical to systematic, to gather insights into a company’s cyber-risk exposure. It has been argued that LLMs can prove pivotal in enhancing a business’s threat detection and response steps, but we must not be beguiled by the infinite possibilities and should instead assess a very complex and real reality.
La breve guía que figura a continuación ofrece una visión general de las razones técnicas por las que los LLM perpetúan sesgos perjudiciales y tóxicos, su impacto en el mundo real y las prácticas responsables para dirigir los LLM hacia un futuro más inclusivo y positivo.
Exploring Hallucinations and Biases in LLMs
Firstly, it’s important to isolate the specific considerations in LLMs and AI, to differentiate them from one another, as the terms can often be used interchangeably.
- AI Hallucinations: AI chatbots like ChatGPT and Bard - among others - can generate falsified content by blending statistics and superfluous ideas in their outputs. The problem lies in the fact that the user can guide the AI to dispense content that aligns with the prompt they provide (even if that’s deliberately or unknowingly made up), and with limitations on its training data, the tool cannot separate fact from fiction intuitively. If not monitored and supervised, hallucinations can manipulate inherently flawed, skewed, and exaggerated data, presenting false narratives that, given humans’ relatively limited understanding of LLMs, can perpetuate discourse that’s easy to take as factual at face value.
- AI Biases: LLMs may be trained on inherently biased training data, which can manifest in numerous ways. Whether that’s spreading stereotypes, prejudices, or phobias, this presents a major ethical and societal concern if such content and data were allowed to be dispensed to the public eye unsupervised and with no oversight. In sensitive and highly regulated sectors like healthcare, news, education, and finance, organizations can ill afford to be perpetuating biased data that only furthers political and social divisions in society.
Understanding How Bias Emerges in AI
AI biases occur in several ways, but some of the most common are highlighted below.
- Biased training data: If the data used to train an AI model contains biases or limited data, those will innately be reflected and projected in any of the tool’s outputs. Text data containing stereotyped or prejudiced portrayals of certain groups, for example, can lead language models to generate biased text with little to no filtering.
- Poor dataset curation: Many datasets contain inadvertent historical biases, many of which lack sufficient diversity, inclusion and equality. Using these datasets without careful curation and balancing can propagate biases that can be harmful.
- Lack of social context: AI systems lack the human social context, experience, and common sense to recognize harmful narratives or discourse. A language model may generate plausible but unethical outputs based solely on pattern matching in a prompt, without inherently understanding the text’s wider and socio-political meaning.
- Lack of transparency: The black-box nature of complex AI models makes it difficult to audit systems for biases. Without transparency into how outputs are generated, biases can slip through undetected. This emphasizes the need for more stringent, regimented, and regular supervision, reviews, and adjustments in AI systems that have already been integrated into business operations.
Given that the global AI market is expected to grow twentyfold by 2030, reaching a valuation of approximately $2 trillion (USD), it’s only fitting that organizations take advantage of the technology ethically and methodically.
Implementing Ethical AI Principles
Formulating high-quality datasets (similarly to what DatologyAI is doing) can, on the surface, be the single biggest and most effective solution in curbing harmful AI text from being generated ad infinitum. However, this is not always practical given how time and resource-intensive data validation and cleansing are. While this should always be a long-term goal, organizations should take simultaneous crucial steps to develop a more responsible, inclusive, and universal AI model within their operations, regardless of sector.
- Establish ethical principles: Develop clear AI guidelines and policies guided by principles of inclusivity, transparency, fairness, accountability and respect for human autonomy.
- Build awareness of harm: Train AI-upskilled teams on the various types of algorithmic harm and mitigation approaches so they can identify issues when using tools autonomously. Human users have a duty to manage and guide AI deployment and continued use in any business context.
- Practice transparency: Openly communicate about data sources, model decision processes, and performance disparities to build trust among all departments and external users. Emphasize the need to highlight biases that some users may have missed or failed to spot, and resolve the matter internally.
- Enable human oversight: Keep all teams in the loop evaluating model outputs before real-world deployment to detect biases. Build feedback loops among users impacted by models to rapidly identify problems and use this to foster a culture of openness and human-first provisions.
- Audit for unfair performance: Continuously test AI models for signs of unfair performance differences across user demographics. Validate any AI-generated outputs regularly and, with thorough scrutiny, assess how fervently users are supervising and guiding content.
Guiding Responsible Generative AI Use
As a powerful emerging category of AI with tremendous potential for misuse, generative AI and LLMs warrant a strong degree of management and oversight in their deployment. Hasty and reactionary deployments of these models into any incumbent business infrastructure and organization, without due diligence on the tools’ validity, will only allow harmful biases to laterally propagate.
However, with ethical and methodical human-first strategies towards any AI system upgrade, organizations can mitigate more of the potential damage and scrutiny that could come their way from consumers, vendors, suppliers, and stakeholders alike. AI’s continued advancement and evolution introduce a plethora of complex risks for businesses, but establishing ethical principles, policies, and safeguards from the outset will allow them to prevent more overt damage from being widely perpetuated.
