The night is dark and full of terrors, the day is bright and lovely and full of hope, as one fairly commercially successful author once put it. It’s a good metaphor for AI, which, like all technology, has advantages and disadvantages.
Stable Diffusion is one of the art-generating algorithms that has inspired amazing creative bursts that have powered apps and even completely original business concepts. However, because it is open source, criminals can use it to produce deep fakes on a large scale, despite artists’ objections that it is making money off of their creations.
What will AI be capable of in 2023? Will regulation be able to stop the worst effects of AI, or are the floodgates already open? Will significant, game-changing new kinds of AI, similar to ChatGPT, impact sectors that were once believed to be immune to automation?
Be prepared for more troublesome AI art-generating apps.
You should anticipate a tonne more similar apps given the success of Lensa, the AI-powered selfie app from Prisma Labs that went viral. Expect them to overly sexualize and change the appearance of women, as well as to be susceptible to being duped into producing NSFW photographs.
Maximilian Gahntz, a senior policy researcher at the Mozilla Foundation, predicted that the consequences of such systems, both positive and negative, would be amplified by the incorporation of generative AI into consumer technology.
For instance, Stable Diffusion received billions of photographs from the internet before “learning” to link particular words and concepts with particular imagery. Text-generating models are frequently amenable to being deceived into promoting offensive ideologies or creating deceptive content.
The Knives and Paintbrushes open research group member Mike Cook concurs with Gahntz that generative AI will continue to be a significant—and problematic—force for change. He believes that generative AI must “finally put its money where its mouth is” around 2023.
For technology to become a permanent part of our lives, Cook stated, “it has to either make someone a lot of money, or have a real impact on the daily lives of the general people.” He continued, “It’s not enough to encourage a group of specialists [to build new tech].” Therefore, I anticipate a strong effort to have generative AI really accomplish one of these two tasks, with varying degrees of success.
The initiative to reject data sets is led by artists.
DeviantArt unveiled an AI art generator adjusted using DeviantArt users’ artwork and based on Stable Diffusion. Long-time users of DeviantArt slammed the platform’s lack of openness in using their uploaded artwork to train the algorithm and expressed their outrage at the art generator.
OpenAI and Stability AI, the two most well-known systems, claim to have made steps to reduce the quantity of damaging information their systems generate. However, it’s obvious that work needs to be done based on many of the generations on social media.
Gahntz compared the procedure to existing disputes over content moderation in social media, saying that the data sets “need active curation to solve these flaws and should be submitted to extensive scrutiny, including from communities who tend to get the short end of the stick.”
In response to criticism from the public, Stability AI, which is mostly sponsoring the development of Stable Diffusion, recently signalled that it will permit artists to opt out of the data set used to train the upcoming Stable Diffusion model. Before training starts in a few weeks, rightsholders will be able to obtain opt-outs through the website HaveIBeenTrained.com.
Unlike Shutterstock, which partners with OpenAI to licence a portion of its picture galleries, Shutterstock does not provide such an opt-out mechanism. But considering the obstacles Stability AI confronts in terms of law and publicity, it’s probably just a matter of time until it follows suit.
It might eventually be forced to act by the courts. A class action complaint filed in the US claims Microsoft, GitHub, and OpenAI of breaking copyright rules by allowing Copilot, GitHub’s programme that intelligently suggests lines of code, to regurgitate portions of licenced code without giving acknowledgement.
GitHub recently added controls to prohibit public code from appearing in Copilot’s suggestions, and it has plans to create a tool that will reference the source of code suggestions—possibly in anticipation of the legal challenge. However, they are faulty measures. The filter setting caused Copilot to generate significant amounts of copyrighted code, including all credit and licence wording, in at least one instance.
As the U.K. considers legislation that would eliminate the requirement that systems trained using public data be used purely non-commercially, expect opposition to increase in the upcoming year.
The use of open source and decentralised initiatives will increase.
A small number of AI businesses, particularly OpenAI and Stability AI, took centre stage in 2022. However, when the potential to create new systems expands outside “resource-rich and strong AI labs,” as Gahntz put it, the pendulum may begin to swing back towards open source in 2023.
According to him, a community-based approach might result in greater scrutiny of systems as they are developed and put into use: “If models are open and if data sets are open, that’ll enable much more of the critical research that’s pointed out many of the flaws and harms associated with generative AI and that’s often far too difficult to conduct.”
Large language models from EleutherAI and BigScience, a project supported by AI firm Hugging Face, are two examples of such community-focused initiatives. The music-generation-focused Harmonai and OpenBioML, a loose collection of biotech experiments, are two groups that Stability AI funds directly.
Although decentralised computing may eventually threaten traditional data centres as open source initiatives mature, money and skills are still needed to build and run advanced AI models.
BigScience has released the open source Petals project, which is a step toward enabling decentralised development. Similar to Folding@home, Petals allows users to contribute their computing power to run complex AI language models that would typically call for a powerful GPU or server.
“Training and using modern generative models requires significant computational investment. According to some rough calculations, the daily cost of ChatGPT is over $3 million, according to Chandra Bhagavatula, a senior research scientist at the Allen Institute for AI. It will be crucial to address this if we want to make this more broadly available and commercially feasible.
Chandra notes, however, that huge labs will continue to have a competitive advantage so long as the procedures and information are kept under lock and key. OpenAI has unveiled Point-E, a model that can create 3D things from scratch when given a text command. Although the model was open sourced, Point-training E’s data’s sources were not disclosed or made available by OpenAI.
For the benefit of more scholars, practitioners, and users, Chandra added, “I do think the open source efforts and decentralisation efforts are very worthwhile. However, despite being open-sourced, many researchers and practitioners still lack access to the best models because of their limited resources.
AI businesses tighten up for upcoming regulations
In the future, regulations like the EU’s AI Act may alter how businesses create and use AI systems. The same may be said for further regional initiatives like New York City’s AI hiring act, which calls for bias audits of AI and algorithm-based tech before usage in hiring, promoting, or recruiting.
Given the increasingly obvious technical shortcomings of generative AI, such as its propensity to spew factually incorrect information, Chandra believes that these regulations are especially vital.
This makes the use of generative AI challenging in many fields where errors might result in exorbitant expenses, such as healthcare. Furthermore, the simplicity of producing false information raises issues with misinformation and deception, she added. In spite of this, AI systems are already making choices that have moral and ethical ramifications.
However, regulation will merely be a threat in the coming year; before anyone is punished or prosecuted, expect a lot more wrangling about laws and legal disputes. However, businesses may still compete for positions in the most favourable categories of new rules, such as the risk categories under the AI Act.
The rule as it stands categorises AI systems into four risk groups, each with different standards and levels of monitoring. Before they are permitted to join the European market, “high-risk” AI systems (such as credit scoring algorithms and robotic surgical apps) must adhere to a set of legal, ethical, and technical requirements. The “minimum or no risk” category of AI (such as spam filters and AI-enabled video games) imposes merely transparency requirements, such as informing consumers that they are dealing with an AI system.
Ph.D. candidate Os Keyes at the University of Washington voiced concern that businesses will pursue the lowest risk level in an effort to reduce their own obligations and visibility to regulators.
Apart from that worry, they claimed that the AI Act was the most advantageous proposal they had seen. “I haven’t noticed much from Congress,” you said.
However, investing is not a guaranteed thing.
Gahntz contends that there is “still a lot of homework remaining” before a corporation releases an AI system broadly, even if it serves the majority of people well enough but severely harms a small number of individuals. “All of this has a business case as well. Customers won’t appreciate your model if it produces a lot of messed-up content, he continued. But certainly justice is also a factor here.
Going into the new year, it’s uncertain if businesses will be convinced by that justification, especially given how eager investors appear to be to invest in generative AI in general.
In the middle of the Stable Diffusion issues, Stability AI raised $101 million from reputable sponsors including Coatue and Lightspeed Venture Partners with a valuation of over $1 billion. As it moves into advanced negotiations to raise further money from Microsoft, OpenAI is reportedly valued at $20 billion. (In 2019, Microsoft made a $1 billion investment in OpenAI.)
According to Crunchbase, the top-performing AI firms in terms of money raised this year were software-based, with the exception of the self-driving firms Cruise, Wayve, and WeRide and the robotics firm MegaRobo. In July, Contentsquare, a company that offers a service that generates recommendations for web content using AI, raised a $600 million investment. In February, $400 million was invested in Uniphore, a company that supplies software for “conversational analytics” (think call centre metrics) and conversational assistants. Highspot, whose AI-powered platform offers real-time and data-driven recommendations to sales people and marketers, raised $248 million in January.
Even though they aren’t as “glamorous” as generative AI, investors may pursue safer wagers like automating the analysis of customer complaints or creating sales leads. That’s not to say there won’t be significant investments that garner notice, but they will be kept for players with influence.
