Episode: Synthetic Reality: How AI is Now Writing the News you thought Humans Wrote Episode Overview For most of the history of the internet, there has been a basic assumption that behind every article, comment, review or opinion piece is a human being; a journalist, blogger, campaigner or just someone expressing a view. That assumption is no longer valid. Today, large volumes of online content are generated automatically, and this is likely to grow. Artificial Intelligence (AI) programs can now produce convincing articles, commentary, social media posts and even videos and images in seconds, resulting in what can be termed as ‘synthetic reality’, an information environment in which it is increasingly difficult to determine whether what you are reading was written by a person or by an algorithm. In this episode, we explore why this has happened and how it has become a common feature of commercial and political discourse. Importantly, it will discuss how listeners may be able to distinguish between the 2 and why it matters. Why is this happening? The introduction of generative AI platforms, such as OpenAI ChatGPT, Anthropic Claude,Microsoft Copilot and others has had a disruptive and transformational impact on how journalism and other content creation is achieved. Initially recognised as highly useful in mitigating much of the time-consuming research, collation and coherence of reference material, through their ability to discover, search and analyse huge datasets and detect patterns, generative AI platforms have enabled journalists and researchers to focus on the more cognitive and value-added elements of content creation. This has enabledorganisations to become more efficient and achieve economies of scale at a time when the way in which media is consumed has significantly changed. An increasing volume of what we read is accessed through digital and social media platforms and traditional news outlets have suffered as a result. Many have therefore embraced the benefits of AI and incorporated them into their workflows to reduce costs and increase output. The latest generation of AI platforms have amplified these advantages, to the extent that many are now able to generate complete and credible articles with minimal human involvement. The utility of generative AI has significantly expanded over its short lifetime, and now includes the following: Scale. AI models are able to generate orders of magnitude more content than a human in a given time.Speed. What would take an experienced human hours or days, can be produced in seconds, by freeing authors and researchers from mundane collection and collation activities, including speech-to-text transcription and translation and searching archives and translating one medium to another.Diversification. Content tailored to diverse consumer groups and even individuals, thereby increasing user engagement and reaching fragmented audiences.Simplifying and contextualising complex source material.‘Knowledge Creation’. The production of text with information content originating in the language model itself, rather than from a source document. Why is AI Content Hard to Identify As Generative Artificial Intelligence (AI) technology continues to evolve, it is becoming increasingly difficult to tell the difference between AI-generated and human-generated content. Advances in core generative algorithms include transformer-based language models that analysehuge quantities of written data to create text that closely mimics human writing, and diffusionmodels that produce highly credible manipulated images. Experiments have demonstrated that humans can distinguish AI-generated text only about half of the time in a setting where random guessing also achieves 50% accuracy. Even following trainingon how to differentiate between the two, or when multiple people work as a team to detect AI-generated text better, detection rates do not improve much. AI detection tools developed specifically to identify AI content often have little more success than humans and are even more challenged when the synthetic content is edited, amended or merged with other data. Normally also created by AI, their effectiveness depends on matching the huge amount of investment in the creation models that utilise huge data sets and computing power, and which areconstantly evolving. The commercial demand for these systems is much lower and consequently they tend to have access to much less data. Traditional AI detection works by sampling and analysing text once it has been written, assessing individual words purely based on the model's probability distribution through searching for statistical patterns that indicates automated production. Another technique, that adopts a completely different approach, is AI ‘Watermarking’. This works by embedding a discrete, machine-detectable signal into content generated by artificial intelligence models to verify ownership and identify that a piece of content was generated by AI. What should we look for? Despite the limitations in detection software, there are still several clues that can lead us to identifying content as AI written or derived. These are only general guides, however, and as generative AI models become more sophisticated, the ability to detect when they have been used is likely to be further impacted: Absence of verifiable attribution. A good place to check first is the authorship. Can you verify the author actually exists, do they have a body of work, are they cited elsewhere?Are references provided at all and if so, do they detail primary sources rather than vague summaries? Can they be found and verified on-line?Language Patterns. AI written content often uses generic language with non-specific phrases such as ‘experts say’ or ‘studies suggest’. It lacks emotional subtlety and tends to use overly formal or complex words or phrases, and sentences are often of the same length.Repetition and Duplication. When researching a given subject, are the same argumentsmade, using the same form of words across multiple sites?Volume vs Depth. A site with lots of articles but little by the way of editorial comment can often be an indicator of AI content farms. Are complex subjects that require personal experience or expert knowledge lacking in analysis and opinion? They may provide definitive statements but without supporting detail or explanation to support the points. Why AI Content Accelerates Misinformation? Synthetic media is proliferating, and here to stay. Given the advantages listed earlier, why should we care? In a word, trust. How can we be sure that what we are reading or seeing, produced by generative AI, is factual and not misinformation? It should be noted that there is a big difference between misinformation, ‘hallucination’ and disinformation, the latter of which will be discussed below. While there is no definitive definition of misinformation, we characterise it as the inadvertent spread of false information without intent to cause harm. While this often produces results that are factually accurate, generative AI models have been known to produce hallucinations, which are answers that have been generated as the probable correct answer based on the patterns in the training data but are in fact incorrect and sometimes totally incoherent. There are numerous circumstances that result in misinformation; for example when a breaking news story is unfolding and before all the details are fully known, or when personal conscious or unconscious bias drives them to an incorrect conclusion or assessment about a given situation. AI doesn’t always make false information more convincing, but it makes it easier to produce in volume. And that matters at a time when information is accessed and consumed in small bites and with limited attention spans. Its sheer persistence and ubiquity mean that misleading content proliferates and reaches a much wider audience. And of course, once information is absorbed, however inaccurate, it is extremely difficult to change that view, and people quickly move on. AI‑driven recommendation algorithms used by platforms can unintentionally: Prioritise sensational or misleading contentIncrease engagement with false narrativesPush misinformation into users’ feeds faster than fact-checked content This creates an environment where misinformation spreads rapidly even without deliberate manipulation. And it can do this in the full spectra of media formats. For example, AI now enables bots (automated accounts on social media) to generate or manipulate text, images, audio and video. This multimodal capability increases the sophistication and believability of misinformation. A 2018 study of Twitter (now known as X) users by researchers at the Massachusetts Institute of Technology found that false information spreads more quickly than accurate information. Popular social media posts, for example, are often easily shared and reposted, without any thought as to the veracity of the content. Once it has gone ‘viral,’ even if the original post is corrected, the false version is likely to endure and proliferate, with no accountability for those doing so. Together, these factors make bots one of the most powerful accelerators of information disorder on social media. The State Actor Angle Unlike misinformation, disinformation is deliberately created and distributed false information that is explicitly designed to...
