Last month, we released Media Critic, an AI tool that can detect bad journalistic practices including attempts to sway public opinion via misleading framing, smear tactics, and other subtle forms of manipulation.
One of the key strengths of Media Critic is that it applies the same consistent set of standards to every article it analyzes regardless of the ideological leanings of the article’s authors. This provides a qualitatively new level of objectivity that was never before possible. Unlike humans who are influenced by various personal biases and tribal allegiances that lead them to hold their adversaries to much a higher standard than their allies, AI systems potentially make it possible to realize the universalist ideal of strictly judging each action by exactly the same unwaveringly consistent ethical standards. The output generated by Media Critic still has plenty of room for improvement, but we are very excited by the potential it has shown so far. Without the technological advances in AI over the last few years, it would have been impossible to produce a tool that can do what it does with anything like the level of sophistication we have achieved.
It helps to have the analyses performed by a non-human entity that has no stake in the outcome, but there are several factors that could still bias the output of Media Critic, which we’d like to draw attention to here.
The evaluation criteria
One potential source of bias is the set of criteria we’ve instructed the AI to use when analyzing articles. Our approach here has been to confine ourselves to journalistic standards that almost everyone appears to agree with. We can all agree that unsourced claims are a problem, that misleading framing is a problem, that sensationalism is a problem, and so on. The criteria are grounded in standard journalistic ethics, and enriched with insights from psychology and linguistics.
These criteria are all outlined in the system prompt used to guide Media Critic’s output. The system prompt also attempts to clarify notions of „objectivity“ and „impartiality“ to steer the AI language model away from naive interpretations of these terms. This is the relevant section from Media Critic system prompt:
Be sophisticated in your discussion of media bias and objectivity. Journalistic objectivity is often naively equated with reporting only the facts of a situation without expressing personal opinions, but there are problems with this definition. While it is desirable for a reporter to stick to the facts when reporting the news, the selective reporting of facts (i.e., the choice of which facts to include and which to exclude) provides a mechanism for introducing bias without the reporter ever needing to express any opinions directly. Thus, a news report that includes only those facts that align with a particular agenda can be extremely biased even though it superficially meets a naive definition of „objectivity“. A truly unbiased report of a news event will encompass all contextually relevant facts that are not already well known. Hence, a more enlightened approach to „objectivity“ is to present „the whole truth and nothing but the truth“.
Note that impartiality is often conflated with objectivity, but the job of a journalist is not to report both sides of an issue with equal weight. If one side says it’s raining and the other says it’s not, the journalist’s job is to look out the window and report what is true. If a journalist is forced to be impartial, it’s generally a sign of inadequate sourcing. Any claims quoted in an article that cannot be easily verified should come with a disclaimer to that effect.
Objectivity is something to aspire to, but transparency about one’s personal biases should also be encouraged.
Also, depending on the type of journalism, it may or may not be appropriate for the author to express personal opinions. Note how the norms of a news article differ from those of an opinion piece, for instance. Regardless of the type of article, a clear demarcation between facts and opinions is always desirable.— Excerpt from Media Critic’s system prompt
The training data
Another potential source of bias, which is unfortunately outside of our control, is baked into the language model itself. To achieve the quality and performance required for Media Critic to function well, we currently use a proprietary language model served by OpenAI. Models like this are trained on vast amounts of text that contains skewed representations and historical prejudices of various kinds that models can potentially internalize during training and then perpetuate.
OpenAI has not disclosed the finer details about what and how data is selected for inclusion in their training sets, but they have provided some very general indications:
ChatGPT and our other services are developed using (1) information that is publicly available on the internet, (2) information that we partner with third parties to access, and (3) information that our users or human trainers and researchers provide or generate. … [W]e do not seek information that we know is behind paywalls or from the “dark web.” We apply filters and remove information that we do not want our models to learn from or output, such as hate speech, adult content, sites that primarily aggregate personal information, and spam.
— OpenAI
There are really two issues here: One is potential biases in the data sources OpenAI uses to train its models, and the other is potential biases in the undisclosed process OpenAI uses to filter this data. We could make a further distinction between biases that are widely condemned by society, and those that are still normalized. As we’ve seen from recent history, what counts as the received wisdom of one generation is often disputed by the next, which has allowed us to make moral progress on issues like racism, sexism, and a host of other issues. Public opinion of wars often shifts on a much shorter timescale in the light of gruesome revelations like the My Lai massacre carried out by US forces in Vietnam, and by exposing the ‚dodgy‘ dossiers and other fabricated intelligence reports that falsely claimed Iraq possessed weapons of mass destruction, claims that were used to justify the US-led invasion of Iraq.
We probably can’t expect language models to be able to completely step outside of the norms of the times we are living in to grasp what future generations will be able to view more objectively.
Fine-tuning
After initial training, a language model undergoes a process of fine-tuning, which helps align it with its role as a helpful AI assistant. This process involves feedback from humans about the quality of its responses, and instilling certain standards of acceptable behavior like not disseminating information that could support criminal activity. It arguably also involves preventing behavior that could attract negative publicity for the company who made it. Hence, commercial pressures could play a role in biasing the behavior of models at the fine-tuning stage, but again, we know very little about the exact policies and processes used by OpenAI or other leading AI tech companies in this regard.
The knowledge cut-off
OpenAI’s GPT-4o model has a knowledge cut-off of October, 2023, which means all the data used to train it predates that. These models can’t learn new things after their training period so GPT-4o doesn’t know who won the 2024 US presidential election, that Kamala Harris replaced Joe Biden as the Democratic candidate in the lead up to it, or that Donald Trump survived an attempted assassination attempt in July.
More of a problem are ongoing news stories that are difficult to understand if you haven’t been reading the news for the last 15 months. The most obvious example is Israel’s onslaught on Gaza. Media Critic knows nothing about Hamas’s attack on October 7th, 2023, nor anything about Israel’s response to it. It doesn’t know that the International Court of Justice has ruled that a plausible genocide is taking place in Gaza, that the International Criminal Court has issued arrest warrants against Israeli Prime Minister Benjamin Netanyahu and former defense minister Yoav Gallant for „crimes against humanity and war crimes“, or that Amnesty International, and Human Rights Watch, and Doctors Without Borders have all independently published extensive reports based on their experience on the ground in Gaza, all of which unambiguously concluded that Israeli forces have indeed been carrying out a genocide. This is a long list of credible institutions that are all saying the same thing and they are far from alone, but Media Critic has none of this context at its disposal, which significantly impedes its ability to assess claims and counterclaims made about this issue. When it analyzes an article that makes reference to a genocide in Gaza, it rightly demands a high standard of evidence be applied to match the seriousness of the accusation, but it demands this for every such article because every time it encounters the accusation, it’s encountering it for the first time unlike readers who are familiar with the evidence because they have been following the story all year.
Media Critic is instructed to explain why it makes the judgments it makes, which should help the user decide whether the basis of a given criticism is valid or not, but we would obviously prefer that it have all the context it needs to speak intelligently about current events. We are exploring ways to provide Media Critic with contextual information that is missing due to knowledge cut-offs, but it will take some time before we can provide a working solution.
While there is still room for improvement in several areas, perfection is an impossible standard. We can only hope to make Media Critic better and better over time. We don’t have any control over bias stemming from how models are trained, but we could potentially find workarounds to deal with the limitations imposed by the knowledge cut-off. In the meantime, it’s important for users to be aware of these issues when using Media Critic.
But how do you think we could improve Media Critic? Leave us a comment.
BTW, you can try out Media Critic here.
0 Kommentare