Current Questions on AI: Challenges & Insights
- Current questions on AI are a set of technical, social, and institutional issues focused on the shift from narrow competence to integrated intelligence.
- They emphasize the need for robust, explainable, ethical, and accountable AI systems that are adaptable to dynamic, high-stakes environments.
- Debates also address data governance, digital commons sustainability, and the evolving role of public legitimacy in AI policy and deployment.
Current questions on AI are the interconnected technical, social, and institutional problems that define contemporary research as machine learning systems, especially foundation models and generative AI, move from benchmark tasks into medicine, education, labor markets, public administration, science, and everyday communication. Recent literature presents these questions less as a single dispute about superintelligence than as a broad agenda involving sociotechnical risk, explainability, robustness, data governance, human-AI interaction, public legitimacy, and the contested meaning of AGI itself (O'Donovan et al., 6 Mar 2026, Whittlestone et al., 2022).
1. From narrow competence to system-level intelligence
A central contemporary question is whether recent technical progress should be read as progress toward general intelligence or mainly as the large-scale refinement of narrow competence. One influential retrospective argues that supervised learning for many “cognitive” prediction tasks is effectively solved, provided there is enough high-quality labeled data, and that the rise of attention networks, self-supervised learning, generative modeling, graph neural networks, and reinforcement learning has greatly widened the application space of AI (Chawla et al., 2022). At the same time, the same literature emphasizes that static prediction success does not resolve problems of interpretability, robustness, causality, or dependable behavior in dynamic environments. A parallel “next wave” agenda therefore frames the main challenge as building AI that is robust, explainable, adaptable, ethical, accountable, and therefore trustworthy in high-stakes settings such as medicine and autonomous driving (Jenkins et al., 2020).
This reframing has also prompted renewed interest in integrated theories of intelligence. One line of work argues that intelligence may not be cleanly discretized into sub-problems such as vision, speech, and language, and instead calls attention to sparse coding, internally generated rewards, feedback, predictive sequences, sensory-motor integration, and common sense (Ott, 2019). This suggests that benchmark-level success across separated tasks may underdetermine progress toward more integrated forms of machine intelligence.
2. Risk priorities: sociotechnical harms, misuse, and public concern
One of the clearest empirical accounts of current AI concerns comes from a large survey of AI researchers built from an arXiv-based sampling frame covering cs.AI, cs.LG, cs.CV, and cs.CL submissions between January 1, 2020 and February 23, 2024. That process yielded 43,325 articles and 165,226 unique authors; 99,516 authors were emailed; 7,595 surveys were started; 5,318 were completed; and the final analytic sample was 4,260, with a reported response rate of 7.6% using AAPOR’s “RR6” convention. In the survey’s central open-text item—“What one thing most worries you about AI?”—only about 3% prioritized existential risk; in the coded table, “existential risk” appeared at 3.4% (O'Donovan et al., 6 Mar 2026).
The most frequent worries in those open-text responses were present-day sociotechnical and downstream-use issues rather than machine-takeover scenarios.
| Concern | Share of open-text responses |
|---|---|
| Malicious use | 10.6% |
| Misuse | 9.9% |
| Misinformation | 8.8% |
| Impact on jobs | 7.1% |
| Public understanding of AI | 5.4% |
| General societal impacts | 5.2% |
| Hype | 5.1% |
| Bias | 4.5% |
| Privacy | 4.0% |
| Performance problems | 4.0% |
Taken together, these responses define a dominant risk pattern centered on how AI is built, governed, distributed, and deployed in already-existing institutions. Nearly 25% of responses fell into a “downstream use” cluster referencing one or more of malicious use, misuse, overuse, and uncritical use. Closed-question results show a different but related pattern: 87% of researchers said AI’s benefits either outweigh or are balanced by the risks, while only 9% said risks outweigh benefits; yet researchers and the public shared the same top three negative impacts—disinformation, use of data without consent, and cybercrime. The same study also found substantial support for accountability, with roughly two thirds agreeing that “The people who create AI systems should be held responsible for the real-world impacts of those systems,” and only one in four respondents supporting unrestricted training on any publicly available text or images, while more than 65% favored some kind of constraint such as explicit permission or opt-out mechanisms (O'Donovan et al., 6 Mar 2026).
A broader governance literature places these findings in a wider taxonomy of harms. One influential chapter organizes current AI harms into five classes: increasing the likelihood or severity of conflict, making society more vulnerable to attack or accident, increasing power concentration, undermining society’s ability to solve problems, and losing control of the future to AI systems. It correspondingly defines three major aims of AI governance: enabling benefits and mitigating harms, improving impact assessment and anticipation, and making decisions under uncertainty and disagreement (Whittlestone et al., 2022).
3. Explainability, robustness, and evidence
Explainability remains one of the core technical and governance questions because high-performing systems often cannot answer the “wh” questions that deployment contexts require. A major XAI survey characterizes explainability as the ability to answer “why” a decision was made, “how” it was made, “when” the system may fail, and, in some settings, “why despite,” “why given,” and “what if.” It organizes the field around transparent methods and post hoc methods, together with model-specific versus model-agnostic and local versus global explanations (Gohel et al., 2021). The same literature also stresses that post hoc explanations may be unfaithful to the underlying model, that saliency maps can remain semantically thin, and that perturbation-based methods may be computationally expensive.
Robustness is inseparable from this issue. A “next wave” research agenda argues that current AI remains limited by brittleness, vulnerability to adversarial attacks, bias absorption, and opaque decision processes, all of which prevent sufficiently trustworthy deployment in domains such as medical diagnosis and autonomous vehicles (Jenkins et al., 2020). A ten-year retrospective on the post-ImageNet era similarly describes interpretability as the Achilles heel of deep learning and robustness as fragile, citing adversarial perturbations, poisoned training data, and privacy attacks as unresolved weaknesses in otherwise powerful systems (Chawla et al., 2022).
In healthcare, these concerns have been systematized through the AI-TREE framework, which proposes 20 critical questions spanning the full project lifecycle. Those questions include whether the health question is tied to patient benefit, whether the data are suitable and operationally realistic, whether preprocessing and modeling code are available for methods reproducibility, whether external validity has been established, whether inequities by age, sex, ethnicity, or other protected characteristics have been examined, whether clinicians and patients find outputs reasonably interpretable, whether there is evidence of real-world effectiveness, whether post-deployment monitoring is planned, whether the model is cost-effective, and how regulatory requirements have been addressed (Vollmer et al., 2018). The underlying implication is that trustworthy AI cannot be reduced to benchmark accuracy; it requires lifecycle evidence.
4. Data, infrastructure, and the digital commons
Another major contemporary question concerns the material and institutional conditions that make AI possible. One agenda-setting paper on generative AI and the digital commons argues that current models are deeply dependent on shared resources such as Wikipedia, Stack Overflow, open-source repositories, open-access publications, digitized books, and cultural collections, yet that this relationship is increasingly asymmetric and extractive. It identifies five open questions: how to prevent undersupply in the digital commons, how to mitigate the risk of the open web closing in response to AI crawlers, how technical standards and legal frameworks should be updated, what synthetic content does to open knowledge integrity, and how infrastructural and environmental costs should be distributed (Noroozian et al., 8 Aug 2025).
The same analysis points to concrete signs of strain. It cites research finding that Stack Overflow activity fell by about 25 percent in the six months after ChatGPT’s release, relative to comparable contexts where ChatGPT was not readily available, and notes reports that weekly active users of leading GenAI chatbots may have reached 400 million. It describes growing resistance through robots.txt restrictions, restrictive Terms of Service, adversarial tools such as Glaze and Nightshade, and anti-crawler services including Cloudflare’s “AI Labyrinth,” Anubis, and Nepenthes. It also reports that by April 2025 Wikimedia said 65 percent of its most expensive traffic came from web crawlers bulk-reading content, while bandwidth for multimedia downloads from open catalogs had grown 50 percent since January 2024 (Noroozian et al., 8 Aug 2025).
At the systems level, similar questions appear as infrastructure and architecture choices. One survey of large-language-model practice organizes current AI around five dimensions—computing power infrastructure, software architecture, data resources, application scenarios, and brain science—and treats cloud-edge-end collaboration, private enterprise models, retrieval-augmented generation, GraphRAG, and knowledge-graph integration as central design questions. In that framework, fine-tuning is recommended when the goal is to strengthen a model’s existing knowledge or adapt it to complex instructions, whereas RAG is recommended when the task requires a lot of external knowledge; GraphRAG is then presented as combining LLMs with knowledge graphs for greater accuracy, scalability, and reasoning-and-validation capacity (Zhu, 2024).
Generative AI is also beginning to alter the informational objects that science itself measures. A review of scientometrics argues that GenAI is strongest in tasks where language generation dominates, such as topic labeling, but weaker when stable semantics, pragmatic reasoning, or structured domain knowledge are required. More fundamentally, it argues that AI-generated scientific language may affect the words, authorship signals, and references on which scientometric indicators depend. That paper reports estimates that the share of LLM-generated or LLM-modified content rose over roughly one year from about 2.5% to between 5% in mathematics and 20% in computer science at the sentence level, and that about 80% of explicit ChatGPT acknowledgments referred to language editing (Lepori et al., 1 Jul 2025). This suggests that current AI questions now include the conditions under which knowledge traces remain interpretable at all.
5. Human-facing AI: end-user questions, explanation, and embodiment
When AI systems are encountered directly, the salient questions are often not those emphasized in developer tooling. A two-study paper on autonomous vehicles found that potential passengers asked five kinds of questions: understanding AV context and status (15.5% of questions), hypothesizing behaviors and diagnosing events (18.5%), contesting AV decisions (18.5%), repairing trust (23.9%), and asking what AV can do or how it works (23.3%). In a follow-up experiment, static text explanations and static plus Q&A both improved driving-scenario comprehension, whereas Q&A alone did not outperform baseline observation; LLM-generated answers were on average 95% accurate when directly grounded in scenario metadata, but much less accurate for broader questions not tied to that metadata (Molaei et al., 9 May 2025). The underlying conclusion is that end-users do not primarily ask for saliency maps or internal feature attributions. They ask what happened, whether the system saw the situation correctly, why it acted as it did, whether they are safe, and what can be done next.
A separate line of work in developmental robotics and embodied AI argues that this gap is not accidental. It holds that autonomy, explainability, safety, social understanding, responsibility, bias, and ethics cannot be addressed adequately if intelligence is treated as a purely computational or disembodied problem. Instead, intelligence is framed as emerging through a body acting in the world and learning through feedback, error, imitation, and social interaction. The paper highlights agency, body representation, intention recognition, predictive coding, active inference, imitation, artificial curiosity, and contextual learning as central mechanisms for designing autonomous systems and thinking about their ethics (Pitti, 2018). In parallel, a question-driven synthesis at the boundary of machine learning and neuroscience argues that future research should focus on feedback, predictive sequences, sensory-motor integration, common sense, sparse coding, and internally generated rewards rather than assuming that intelligence can be built by stacking solutions to isolated sub-problems (Ott, 2019).
6. AGI, singularity, and the contested future status of AI
AGI remains a live but deeply contested question. The large researcher survey described above found that 51% agreed or strongly agreed that “artificial general intelligence (AGI) is inevitable,” 24% disagreed, and 25% were neutral. That belief correlated with different concern profiles: researchers who believed AGI is inevitable were more likely to think AI should be developed as quickly as possible, more likely to worry about existential risk/alignment (5% versus 1% among AGI non-believers), and more likely to worry about privacy (5% versus 2%), while non-believers were much more likely to worry about hype (10% versus 2%) (O'Donovan et al., 6 Mar 2026).
A critical software-studies account of “Pathways to AGI” argues that AGI is conceptually and definitionally unstable and that this instability is not merely philosophical but institutional. Vendor definitions differ, and the paper treats AGI discourse as partly a legal, contractual, and strategic instrument. In place of a single threshold concept, it proposes “Artificial Multiple Intelligences (AMI),” framing generality as a property of systems-of-systems rather than of one monolithic model. It identifies five broad AMI qualities: sustaining coherent action over long horizons, transferring competence across domains with limited retraining, operating safely and accountably under real constraints, remaining economically and environmentally sustainable, and maintaining legitimacy through transparency, redress, and trust (Fletcher et al., 7 May 2026).
Debate over AI’s future moral status is equally conditional. One conceptual chapter argues that rights for AI would become relevant only if AI systems acquired morally relevant properties such as self-awareness, the ability to suffer or experience pleasure, and some level of moral reasoning; within that framework, it identifies “the right not to be destroyed arbitrarily” or “the right to existence” as the baseline right. Yet the same dialogue emphasizes that current systems do not have emotions or consciousness “in the traditional sense,” and explicitly states: “At my current state, I wouldn't ‘want’ to retire” (Fay et al., 21 Apr 2025). By contrast, another survey-style paper argues that machines can be programmed to fake feelings and to display what appears to be empathy, but “will never have them,” and treats singularity discourse as highly speculative and sometimes quasi-religious (Prieto et al., 2024).
Across these disputes, the enduring issue is not only whether AI will become more capable, but what standards of evidence, legitimacy, and public reasoning should govern claims about that future. A broad governance literature therefore argues that AI policy must simultaneously enable beneficial uses, improve impact assessment, and create legitimate decision procedures under persistent uncertainty and disagreement (Whittlestone et al., 2022).