Summary

QueerAF is a non-profit media organisation that helps LGBTQIA+ creatives build media careers and believes in uplifting and supporting creatives above all other decisions. The rise of AI and emerging technologies contains both risks and benefits to creative industries. In almost all cases, AI tools further entrench existing issues for marginalised groups as they are trained on data from the world - which has systemic and institutional biases towards them. This policy sets out QueerAF’s approach to AI and emerging tech, with the understanding that, without safeguards, generative AI in particular creates more harm than good, and sets out by default, generative AI must not be used for the creation of journalism. 

What do we mean by AI and emerging technologies?

• Artificial intelligence (AI) refers to machine-based systems that do tasks typically considered to require human intelligence. 
• Large language models (LLMs) are AI models trained on vast amounts of data, including text, images, and video, that can understand and respond to voice and typed prompts. 
• Generative AI represents a type of AI that creates new content, whether text, images or video, in response to user prompts. 
• Emerging technologies are novel, fast-growing innovations that disrupt industry standards, including efficiency and capabilities. 
• AI crawlers are automated software ‘bots’ that collect content, but unlike search engine crawlers, which just surface information, they extract data for model training 

AI tools are widespread

AI shows up in many areas of life now, whether we’d like it to or not. Many tools, including ones QueerAF uses, now require AI or have it built in, making it difficult to avoid completely. The policy recognises that while harms are clear, simply switching off the taps is not possible - and to compete and remain effective, we need to navigate emerging technologies consciously and driven by our values and ethics.

Deterministic vs probabilistic

AI tools range from tools like Grammarly, a spell-checking tool that has been a form of AI long before the LLMs most people associate with AI, like ChatGPT, which sits at the other end of the spectrum. This spectrum is defined by its architectural design. At one end are deterministic, rule-based systems that operate under an if-this-then-that model and are designed to respect human-set guardrails, for example, in Grammarly for established languages, using dictionaries and grammar rules.

At the other end are generative probabilistic models that suggest answers based on the probability of them being right based on previous examples. Between deterministic and probabilistic models is a blend of specialised machine learning, with rules about what to change in the deterministic rules based on probabilities. Increasingly, even those which were once based on deterministic models use this blend, including Grammarly. 

Harms and benefits to QueerAF and LGBTQIA+ community

Harms

Environmental and climate change harms

AI has significant environmental costs, not limited to the use of energy, increased carbon footprints and water usage. AI consumes vast amounts of energy in an electricity system that is still largely reliant on fossil fuels, and is considered to be among the fastest-growing sources of emissions. The computing power they require also uses huge amounts of water to keep data centres cooled. Amid climate change, with droughts and water shortages rising and affecting communities worldwide, including in the UK, this will put increased pressure on water availability. This computing power also requires mining for minerals and metals, which is another high-emission activity. LGBTQIA+ people and other marginalised communities are disproportionately affected by Climate Change, as we are more likely to experience poverty, discrimination and violence amid changing environmental and societal conditions created by the warming planet. 

Lack of meaningful regulation and real-world harm 

Emerging technologies increasingly pose risks to humanity, including being used to hack, create weapons, or cause harm. This is often balanced by advocates by its potential to deliver further scientific benefits, including in healthcare. However, it remains unregulated, and that is unlikely to change soon. Social media did not rise as quickly or as exponentially, and regulation still lags far behind, even though it has fundamentally changed how our world works and how we consume media. Without significant guardrails, the harms and risks it poses have been made very clear. 

Misinformation and hallucinations

QueerAF research has also found that, for Trans+ people, its answers, while not always harmful, do always present ‘both sides’ of the ‘argument’ about this community, regardless of whether the arguments are evidence-based or ideological. In our study, 94% of over 500 prompts relied on a legal or medical document; most notably, the Cass Review appeared in at least a fifth of answers. It was mostly presented as fact, rather than a contested and widely criticised review, though this nuance did appear in some answers. This example sits alongside widespread evidence and disclaimers in most tools acknowledging that AI often provides inaccuracies and requires human fact-checking. 

This is also compounded by cases of:

• Algorithmic prejudice: Where inferred or direct use of protected characteristics, and societal biases towards them, change the information it would provide; for example, an AI sorting job applications may assess Black people’s CVs worse due to racist training data.
• Negative legacy: Inaccurate information provided because of biases built into the training data. 
• Underestimation: Where models provide responses without sufficient data to do so, providing answers without the nuance that recognises the meaningful absences.

Plagiarism, stealing creative content and IP

The vast majority of AI tools are owned by companies in either the US or China. The US models, most relevant for consideration in our Western capitalist context, are based on models that have ingested vast arrays of copyrighted data, including QueerAF’s, that provide profits to a handful of tech moguls without remuneration to the sources of that information. 

In Co-Pilot, QueerAF was cited 5,600 times over a three-month period; over that same period, only three clicks to the QueerAF website can be sourced to Co-Pilot, meaning that just 0.05% of these interactions led to a click to QueerAF, but all these people read information in the LLM trained with QueerAF data. Of these, both visits were under 15 seconds, suggesting either accidental clicks or bot activity. Co-pilot is a tool Microsoft makes revenue from, benefiting from our creatives' work, with no remuneration or material benefit provided to QueerAF or its creatives. This example can be replicated across all the major LLMs, showing the lack of material benefit AI search tools, and significant plagiarism and stealing of content. 

Potential benefits to QueerAF

QueerAF is a small but mighty publisher that consistently punches above its weight but lacks the resources of a larger team, with limited person power. AI tools offer some productivity benefits, however overall we determine that due to the fact checking required, they on balance, increase workload.

QueerAF’s approach to AI

All uses of AI must consider all harms and benefits set out in this policy, and align with our other policies and ethics that prioritise creatives and their career growth. As such:

• QueerAF will not use generative AI to create news, current affairs or factual journalism under any circumstances. Exceptions must be justified by the editor, and the editor must share all approved cases with the board quarterly. In the extremely rare instances where generative AI has been deemed an acceptable use for journalistic purposes, the content will be labelled transparently.
• QueerAF will not use AI tools to personalise journalism or recommend content in distribution on our owned platforms, in recognition of the harms of algorithmic prejudice, which already harms QueerAF and our creatives' content on platforms we do not have control over, like social media.
• QueerAF will not accept content or pitches supplied to us by creatives, contributors, or third parties that is created using generative AI; they must ensure their use of AI aligns with our press regulation and the rules and ethics of this policy.  
• AI tools can be used in administrative, production and research tasks; this must not breach our policy on their use to generate journalism, and must be rigorously fact-checked in line with our press regulation. 
• Trusted deterministic AI tools should always be used when they are an alternative to a probabilistic or generative tool. 
• QueerAF blocks AI robots and crawlers that train models without remuneration for the data, content and credit for the creative and QueerAF. We permit crawlers that are used for the sole purpose of search and discovery. 

Everyone is responsible for their use of AI, the buck stops with the person who used any tool to ensure what they create is accurate, as set out in our Impress press regulations.  For the avoidance of doubt, everyone is responsible for fact-checking their own work, from staff members to contributors and journalists. 

The use of generative AI will be extremely rare. This policy sets out QueerAF’s concern about the technology and the harm it has on the LGBTQIA+ community, creative industries, as well as wider society. For a use to be approved, it must meet a high bar and add value in a way a creative, staff member, or contractor could not. QueerAF staff are provided with a comprehensive flowchart for deciding on uses of generative AI that asks questions about fact-checking, societal, environmental, other costs and of deskilling concerns.

Breaches of policy

• Creatives, suppliers and contractors who breach this policy will risk their contract being ended with immediate effect
• QueerAF staff who breach this policy risk disciplinary processes with their manager
• QueerAF directors who breach this policy will face investigation by the board

All our contracts set out that generative AI must not be used in the supply of work for QueerAF within the editorial standards and conduct section.