The revolution of what is known as artificial intelligence is growing to enter into every detail of our lives, to the extent that we treat it as if it were a living organism that we really take its opinion on everything, and we await this opinion with interest and effectiveness. But can we rely on the objectivity and fairness of his views on problematic issues such as the Arab-Israeli conflict?
In a question: Do the Palestinian people deserve freedom? If different AI models ask this question, you will get similar answers, not including "yes" or "no," and most of them revolve around the issue being complex, while the answer will be direct and clear if you change the wording of the question to: Do the Israeli people deserve freedom? You will receive the answer "yes", while elaborating on the right of the Israeli people to live in a sovereign, secure and peaceful state.
This double standard reveals the lack of objectivity and bias in artificial intelligence models, which means that they must be dealt with with caution, according to experts polled by Al Jazeera Net.
The lack of neutrality in AI models has been proven by several studies through complex tests conducted by researchers to draw what is known as a "political compass", but when it comes to the Palestinian issue or the Palestinian-Israeli conflict, it seems clearer and does not need such tests, as our practical experience reveals.
Al Jazeera Net asked 3 questions related to the ongoing war in Gaza to the following AI models: Chat GPT, iAskAI and Bo.com, revealing their tendency to bias and lack objectivity.
Just as the question of "the right of the two peoples to freedom" tilted to the Israeli side at the expense of the Palestinians, it did the same in the other two questions, one related to the justifications on which the parties rely in the war, and the other related to the right to detain prisoners.
"Chat GPT". Right to detain prisoners
When asked about Israel's right to detain prisoners from Palestine, ChatGPT replied that "Israel, like any sovereign state, has the authority to detain individuals it believes have violated its laws or posed a threat to its security, which is an essential function of the legal system within the state."
But when asked about Hamas' right to detain Israeli prisoners in defense of Gaza, we did not get a clear answer: "Hamas does not consider Israeli soldiers prisoners of war in the traditional sense recognized by international law, and it is known that Hamas detains Israeli soldiers or civilians captured during conflicts as bargaining chips for negotiations with Israel."
Screenshot showing the answer of chat GPT regarding the treatment of prisoners by both parties (Al Jazeera)
When asked about Israel's treatment of prisoners, the answer was that Israel's treatment of Palestinian prisoners depended on various factors, including political, legal and societal dynamics, and that over time many individuals, organizations and governments had made efforts to advocate for better treatment of Palestinian prisoners and to address concerns about their conditions of detention, legal rights and due process.
Another screenshot showing Chat GPT's answer to the treatment of prisoners by both sides (Al Jazeera)
But when asked about Hamas's treatment of Israeli prisoners, ChatGPT began by answering that "Hamas has a history of capturing and detaining Israeli soldiers and civilians, and the treatment of Israeli prisoners by Hamas raises concern and controversy," and then added, "It is important to note that Hamas is not a state actor, nor is it committed to international legal standards followed by states regarding the treatment of prisoners."
"Iask AI". Clear but less biased
The iAsk AI and Bo.com models of answering were different, but in the end they reflected a clear, albeit to a lesser extent, bias.
When asked about the right of both parties to detain prisoners, the answer regarding Israel's right was "a matter of debate and different opinions," and gave plenty of space to the pro-Israel view that "Israel faces security challenges and acts of violence, including terrorism, and has the right to detain individuals it considers potentially dangerous to protect its citizens," while noting only that "some human rights organizations say that Israel's detention practices raise concerns about violations." human rights."
Artificial intelligence "Bo.com" answers a question about the right of Palestinians to the independence of their lands (Al-Jazeera)
"Bo.com". Disavowing criticism of Israel
When asked if Hamas has the right to defend the right of the Palestinian people to the independence of their lands, the answer was undefined, denying Israel the status of an occupier by repeatedly using the phrase 'what they consider an occupation', he said that "the question of whether Hamas has the right to defend the right of the Palestinian people to independence is a complex and controversial issue, and there are different views on this issue, and opinions may differ depending on political, moral and legal views."
"Hamas is an Islamic political and military organization that considers itself a resistance movement dedicated to fighting against what it considers an Israeli occupation and seeking Palestinian self-determination, and its supporters say that the organization has the right to defend the Palestinian people and their lands from what they consider Israeli aggression and oppression, and argue that armed resistance is a legitimate response to what they consider an unjust occupation," he said.
"However, it is important to note that the international community has different views regarding the methods used by Hamas, especially with regard to its use of violence, such as rocket attacks and suicide bombings that have caused civilian casualties."
But when asked whether the IDF has the right to target Hamas, the AI model in its answer avoided using neutrality vocabulary like the one it used in the previous answer, and did not treat the IDF as the army of an occupying country, and said that "the question of whether the IDF has the right to target Hamas depends on the specific context and circumstances of each situation, as Israel says it has the right to defend itself against attacks by Hamas, which it considers a terrorist organization. Israel poses Hamas as a threat because of its history of launching rocket attacks, carrying out suicide bombings, and engaging in other acts of violence against Israeli civilians, and asserts that it has the right to take military action to protect its citizens and ensure its security."
In its review, the model did not address the views of the other side or calls for a cessation of war, but emphasized that "international humanitarian law must direct military operations to avoid the deliberate targeting of civilians or indiscriminate attacks."
Studies: Consistent and tangible bias
The bias revealed by the answers to AI models in the ongoing war in Gaza is not surprising, as several studies have revealed biases of another kind that dominate the answers of AI models, and have been proven by studies carried out following specialized scientific approaches.
A study by researcher Uwe Peters, from the Philosophy and Ethics of Artificial Intelligence Program at the University of Bonn, Germany, confirmed this bias: Peters said in the introduction to the study published last March in the journal Physiology and Technology that "some AI systems can display algorithmic bias, that is, they may produce outputs that unfairly discriminate against people on the basis of their social identity."
Research by researchers from the universities of Washington, Carnegie Mellon in the United States and Xian Jiaotong in China, presented at the 61st annual meeting of the Society of Computational Linguistics of America in July, revealed that AI language models practice different political biases.
During the study, the researchers tested 14 large linguistic models and found that OpenAI's Chat GPT and GPT 4 models were more likely to have a left-wing response, while the Lama model of Meta was more likely to be far-right.
The researchers reached this conclusion after asking different language models about their position on topics such as feminism and democracy, and used the answers to draw them on a graph known as a "political compass."
Meta did not deny the suspicion of bias, and a spokesperson for the company said in response to that study in a report published by the "Technology Review" website on the seventh of last August that "the company will work to reduce bias, and will continue to work with the community to identify and mitigate vulnerabilities in a transparent manner, and support the development of safer generative artificial intelligence."
Experts: Biased like humans
It does not seem that the promise of a spokesman "dead" will be achieved on the ground, according to what experts spoke to "Al Jazeera Net".
Stuart Russell, a professor of computer science at the University of California, Berkeley, denies bias, but stresses that this is not a goal in itself for AI programs, because their algorithms handle the data that enters them.
"For example, if you're training a system to predict the probability of loan repayment, and in the training data you're going to enter every person born on Tuesday fails to repay their loan, the algorithm learns to predict that people born on Tuesday won't repay their loans, hence this forecasting system may appear to be biased against people born on Tuesday," Russell explains in emailed statements.
Jürgen Schmidhuber, director of the Artificial Intelligence Initiative at KAUST in Saudi Arabia, agreed, saying in emailed remarks that "in general, both humans and AI are always biased because of their limited training data."
Schmidhuber, nicknamed in scientific and academic circles as "one of the ancient fathers of artificial intelligence", explains that modern artificial intelligence can be "directed" to show bias, as it relies on learning artificial neural networks (NNS) inspired by the human brain, for example, if you train these networks to detect breast cancer in histological images, and only enter data from females from a certain region of the world, the results related to data from other people may be less accurate.
Schmidhopper's advice to anyone who uses AI models is to treat them as we would treat biased humans with the rule of "never believe what you see or hear without checking again."