Training such models is a data-intensive task; the more and better the data, the higher the quality of the model and its outputs. Interestingly, finely tuned models trained on high-quality, specific data are starting to perform comparably to larger models like ChatGPT.

AI’s utility extends across a wide array of industries, including coding, design and advertising. One startup is employing generative AI to augment the efficiency of call centres by processing accumulated data. Other applications include building training databases for smaller language models. 

Despite the promising possibilities, AI also has two main drawbacks that make it a potent risk for disinformation. The first is that it is susceptible to ‘hallucinations’, generating false information inadvertently. The quality and nature of the data used to train the model significantly influence its outputs, and this principle applies irrespective of the application. The second drawback is the intentional use of generative AI to create false information, like deepfake photos of an explosion at the Pentagon, and distribute them as propaganda. AI’s images and text are very convincing, and whilst they can fool a lot of people now, these models are not yet perfected and it’s possible to spot inconsistencies in images and text, like how the fence around the Pentagon melds into the crowd barriers. However, whilst people can spot these inconsistencies now, as the models are perfected, this won’t always be the case. 

AI also carries substantial legal and ethical concerns. It is challenging to guarantee the outcomes of AI products due to their inherent unpredictability. Additionally, Kitty explained how “AI models may harbour biases, as evidenced by GPT’s politically left-leaning outputs.” The consequences of this could be particularly severe when AI is being used in biomedicine. There are fears that if, for instance, AI is not fed enough data from patients from minority ethnic groups that it will struggle to diagnose them accurately. There has already been a precedent for this with facial recognition software which has historically struggled to distinguish facial features amongst non-white ethnicity groups. Companies also run the risk of potential lawsuits if their AI models inadvertently infringe copyrights, as in the ongoing court case between Getty Images and generative photo AI, Stable Diffusion which had generated “new” images with Getty watermarks, revealing that its training data included copyrighted Getty images. 

AI regulation has become a hot topic, with no solid approach laid out even after a Senate inquiry involving US AI company leaders like OpenAI’s Sam Altman. However, one of the key recommendations from this inquiry was that the data sets used to train AI models should be made transparent. AI companies can currently keep their precise training data a commercial secret, but knowing what a model has been trained on makes it much easier to understand why it has given certain outputs – and therefore provides something to trace back and potentially regulate. The EU is writing a bill at the moment that Sam Altman says might cause OpenAI to cease operations in Europe, as it demands transparency over ChatGPT’s training data. Meanwhile, it’s expected that US regulation will go in a different direction, so we may see regulatory divergence across the world, explained Kitty.