Artificial intelligence: Difference between revisions
m Added reference archive. Tags: Mobile edit Mobile web edit Visual edit |
section about system-prompts and lack of transparency |
||
| (5 intermediate revisions by 3 users not shown) | |||
| Line 7: | Line 7: | ||
[[wikipedia:Generative artificial intelligence|Generative artificial intelligence]] models are trained through vast amounts of existing human-generated content. LLMs gather statistics on word patterns, which allows the model to generate sequences of words that seem similar to what a person might have written. However, an LLM does not understand anything; they cannot reason. They generate randomly modulated pattern of tokens. In this way, they function similarly to autocomplete. | [[wikipedia:Generative artificial intelligence|Generative artificial intelligence]] models are trained through vast amounts of existing human-generated content. LLMs gather statistics on word patterns, which allows the model to generate sequences of words that seem similar to what a person might have written. However, an LLM does not understand anything; they cannot reason. They generate randomly modulated pattern of tokens. In this way, they function similarly to autocomplete. | ||
People reading sequences of tokens sometimes perceive things they think are true. Sequences that do not make sense to the reader, or that are false, are called [[wikipedia:Hallucination (artificial intelligence)|hallucinations]]. LLMs are typically trained to produce output that is pleasing to people, exhibiting [[dark patterns]]. For example, they produce output which seems confidently written, use patterns which praise the user (sycophancy), and employ emotionally manipulative language. | People reading sequences of tokens sometimes perceive things they think are true. Sequences that do not make sense to the reader, or that are false, are called [[wikipedia:Hallucination (artificial intelligence)|hallucinations]]. LLMs are typically trained to produce output that is pleasing to people, exhibiting [[Dark pattern|dark patterns]]. For example, they produce output which seems confidently written, use patterns which praise the user (sycophancy), and employ emotionally manipulative language. | ||
People are accustomed to interacting with others, and many overestimate the abilities of things that exhibit complex, person-like patterns. Promoters of “AI” systems take advantage of this tendency, using suggestive names (like “reasoning” and “learning”) and grand claims (“PhD level”), which make it harder for people to understand these systems. | People are accustomed to interacting with others, and many overestimate the abilities of things that exhibit complex, person-like patterns. Promoters of “AI” systems take advantage of this tendency, using suggestive names (like “reasoning” and “learning”) and grand claims (“PhD level”), which make it harder for people to understand these systems. | ||
| Line 27: | Line 27: | ||
There are several concerns with using online AI models like [[ChatGPT]], not only because they are proprietary, but also because there is no guarantee of where your data will be stored or used. Recent developments in local AI models offer an alternative to online AI models, which can be downloaded from platforms like [https://huggingface.co/ HuggingFace] and used offline. Common models to run include Llama ([[Meta]]), DeepSeek ([[DeepSeek]]), Phi ([[Microsoft]]), Mistral ([[Mistral AI]]), Gemma ([[Google]]). | There are several concerns with using online AI models like [[ChatGPT]], not only because they are proprietary, but also because there is no guarantee of where your data will be stored or used. Recent developments in local AI models offer an alternative to online AI models, which can be downloaded from platforms like [https://huggingface.co/ HuggingFace] and used offline. Common models to run include Llama ([[Meta]]), DeepSeek ([[DeepSeek]]), Phi ([[Microsoft]]), Mistral ([[Mistral AI]]), Gemma ([[Google]]). | ||
In some cases, AI models can be hijacked for malicious purposes. Demonstrated with Comet ([[Perplexity]]), users can run arbitrary prompts to the browser's built-in AI assistant by hiding text in the HTML comments, non-visible webpage text, or simple comments on a webpage.<ref name=":0">{{Cite web |date=Aug 20, 2025 |title=Tweet from Brave |url=https:// | In some cases, AI models can be hijacked for malicious purposes. Demonstrated with Comet ([[Perplexity]]), users can run arbitrary prompts to the browser's built-in AI assistant by hiding text in the HTML comments, non-visible webpage text, or simple comments on a webpage.<ref name=":0">{{Cite web |date=Aug 20, 2025 |title=Tweet from Brave |url=https://nitter.us.catsarch.com/brave/status/1958152314914508893 |url-status=live |archive-url=https://web.archive.org/web/20260321120531/https://nitter.us.catsarch.com/brave/status/1958152314914508893 |archive-date=21 Mar 2026 |access-date=Aug 24, 2025 |website=X (formerly [[Twitter]])}}</ref> These arbitrary prompts can then be exploited to obtain sensitive information or gain unauthorized access to high-value accounts, such as those for banking or gaming libraries.<ref>{{Cite web |date=Aug 23, 2025 |title=Tweet from zack (in SF) |url=https://nitter.us.catsarch.com/zack_overflow/status/1959308058200551721 |url-status=live |archive-url=https://web.archive.org/web/20260321120841/https://nitter.us.catsarch.com/zack_overflow/status/1959308058200551721 |archive-date=21 Mar 2026 |access-date=Aug 24, 2025 |website=X (formerly [[Twitter]])}}</ref> See [[wikipedia:Prompt_injection|Prompt injection]]. | ||
===Unethical maintenance of data centers=== | ===Unethical maintenance of data centers=== | ||
Due to heavy investments into and increased use of generative AI and LLMs, many data centers have been constructed to host LLMs. These data centers consume large amounts of power and water, in order to power and cool the computer systems running the models. Residents that live in cities where AI data centers have been constructed have complained of an increase in their electricity bills despite no change in their personal usage.<sup>[<nowiki/>[[Consumer Rights Wiki:Verifiability|citation needed]]]</sup> According to a research video by Benn Jordan, these data centers (as well as fracking operations and natural occurrences) cause a high amount of sound pollution, which can cause various symptoms.<ref> https://www.youtube.com/watch?v=_bP80DEAbuo ([https://preservetube.com/watch?v=_bP80DEAbuo Archived])</ref> | Due to heavy investments into and increased use of generative AI and LLMs, many data centers have been constructed to host LLMs. These data centers consume large amounts of power and water, in order to power and cool the computer systems running the models. Residents that live in cities where AI data centers have been constructed have complained of an increase in their electricity bills despite no change in their personal usage.<sup>[<nowiki/>[[Consumer Rights Wiki:Verifiability|citation needed]]]</sup> According to a research video by Benn Jordan, these data centers (as well as fracking operations and natural occurrences) cause a high amount of sound pollution, which can cause various symptoms.<ref> https://www.youtube.com/watch?v=_bP80DEAbuo ([https://preservetube.com/watch?v=_bP80DEAbuo Archived])</ref> | ||
=== Hidden directives === | |||
Most AI apps include an initial "root"/"system" prompt given to the AI, which is hidden from the user. Some corporations go to great lengths to keep those prompts hidden, and to avoid leaking it to the user. Some projects attempt to bring back transparency to these tools, in spite of the restrictions.<ref>https://github.com/elder-plinius/CL4R1T4S</ref> | |||
==Further reading== | ==Further reading== | ||
*[[Automatic content recognition]] | *[[Automatic content recognition]] | ||
==External links== | |||
*[https://aisafety.dance/ Nicky Case, ''“AI Safety for Fleshy Humans”'', Hack Club (2024)] | |||
==References== | ==References== | ||
{{Reflist}} | {{Reflist}} | ||
[[Category:Artificial intelligence]] | [[Category:Artificial intelligence]] | ||