How to Efficiently Collaborate for a Research Project in the Era of AI?

Does this sound familiar? At the start of a collaboration, you hit it off with your partner, feeling like you’ve found a “lifelong collaborator.” You brainstorm directions, datasets, future papers, and naturally assume this will blossom into a long-term professional relationship.

But once the real work begins, things slowly shift: promised results are repeatedly delayed, agreed-upon roles are compromised, and communication dwindles. It’s only at critical milestones, like deciding paper authorship, patent ownership, or wrapping up the project that both sides suddenly realize they have fundamentally different definitions of “collaboration.” Rest assured, it’s not just in your head. Many scientific collaborations fail not due to technical hurdles, but because of unstated expectations, unrecognized contributions, and the absence of a clear exit strategy.

At its core, scientific collaboration is more than just an exchange of knowledge; it’s an ongoing negotiation of boundaries, contributions, trust, and the allocation of credit. This article moves beyond the simple question of “whether to collaborate” to address a more pragmatic issue: given that testing boundaries and power dynamics are natural parts of any partnership, how do we protect ourselves while successfully and pleasantly getting the job done?

1. Why Do Collaborations Turn Into Mutual Boundary Testing?

Ideally, scientific collaboration is complementary: you provide the methods, I provide the data; you offer the theory, I offer the practical application; you excel at experiments, while I shine at data analysis and writing. Everyone leverages their strengths to boost overall productivity.

Reality, however, is rarely so straightforward. Many collaborations lack a clear agreement, relying instead on a vague verbal commitment like, “Let’s work on this together.” This is precisely where the problem begins. As the collaboration unfolds, partners constantly test the waters, essentially trying to gauge what the other person is willing to tolerate.

1.1 The Core of Boundary Testing: Balancing Interests Amidst Ambiguity

A common pitfall in interdisciplinary and cross-institutional work is poorly defined boundaries. Researchers from different backgrounds often have varying definitions of a “reasonable contribution,” “normal communication,” and “fair compensation.” This ambiguity manifests as specific boundary-testing behaviors during the project:

Probing Type	Typical Scenario	Hidden Intention
Testing Timelines	Promising results “next week,” but failing to deliver or follow up.	Seeing if you will chase them down and how long you will wait.
Testing Effort	Committing to equal work, but ultimately only wanting a co-author credit or offering occasional feedback.	Seeing if you will step up to do the heavy lifting.
Testing Resources	Using your data, lab equipment, or students’ time without offering clear credit or compensation.	Seeing if you will enforce your professional boundaries.
Testing Authorship	Avoiding discussions about author order early on, then making new demands right before submission.	Seeing if you will cave under pressure.

These behaviors aren’t always malicious; some people are genuinely overwhelmed, and others simply lack the self-awareness to realize they are draining their partners. Regardless of intent, these tests constantly shift the power dynamic: the person who is more anxious, more reliant on the project, or too polite to speak up inevitably ends up at a disadvantage.

This testing occurs because both sides are assessing risk. Unsure if the other party is reliable or what the return on investment will be, people use subtle moves to test the waters.

A truly mature collaboration doesn’t eliminate these tests entirely; instead, it brings them out into the open: When are the deliverables due? Who is handling what? How will contributions be tracked? How will credit be distributed? Clarifying these details early on prevents friction down the road.

1.2 Flashpoints: When Conflicts of Interest Erupt

Issues rarely surface on day one. At the outset, everyone is usually polite and eager to please. The real problems tend to emerge as the project hits critical milestones.

The Mid-Point: Once the honeymoon phase ends, mismatched expectations become obvious. This is when partners often realize things aren’t going as envisioned. For instance, Person A expects Person B to handle the full data analysis, while B thought they were only consulting on statistics; A expects weekly check-ins, while B prefers to communicate only when there are final results. These discrepancies don’t mean someone is “wrong”, but they simply highlight a failure to align on task boundaries.
The Pre-Publication Stage: Papers, patents, and awards act as magnifying glasses for unresolved conflicts. This is the most sensitive phase of any partnership. Decisions about first authorship, corresponding authorship, future data rights, and patent revenue distribution directly impact personal performance reviews, promotions, and funding. If these weren’t discussed early on, bringing them up now feels like “settling scores.” Both sides feel they made vital contributions, leading to mutual resentment.
Project Wrap-up: How resources and future collaborations are handled determines whether trust survives. Wrapping up isn’t just about finishing; it’s about reallocating resources. Who controls the final outputs? Can the data be reused? Who leads the next paper? How is IP managed in corporate partnerships? These answers dictate whether a collaboration ends gracefully or leaves a bitter taste.

These conflicts are highly sensitive because the academic evaluation system is inherently competitive. Authorship order, grant allocations, and patent rights might seem like minor administrative details, but they dictate career trajectories.

Therefore, rather than accusing each other of being “unreasonable” after the fact, it’s better to acknowledge reality upfront: scientific collaborations involve tangible stakes, and those stakes require serious, proactive discussion.

2. Recognizing the Differences Among Collaborative Partners

Many collaborative breakdowns occur not because a partner is “bad,” but because you’re applying the rules of one type of relationship to another. Collaborating with labmates, external institutions, and industry partners requires entirely different playbooks: labmate collaborations rely on familiarity, cross-institutional collaborations rely on contractual agreements, and industry-academia collaborations rely on aligned goals. Different partners carry entirely different risks.

2.1 Collaborating with Labmates: The Hidden Dangers of Familiarity

Teaming up with peers in the same research group or under the same advisor feels like the path of least resistance: you see each other daily, share similar research interests, and use the same data and methodologies. These collaborations boast low communication barriers, making them easy to kick off for several reasons:

Mutual understanding: You already know each other’s research focus, skill limits, and work habits.
High Efficent communication. You are in the same lab, the face-to-face communication is more efficent than remote.
Informal problem-solving: Issues can be hashed out during lab meetings, in the office, or over casual chats.
Advisor mediation: When disputes arise, the Principal Investigator (PI) or lab rules act as a buffer.
Shared assets: Existing datasets, codebases, lab protocols, and writing templates are easily repurposed.

However, these partnerships are riddled with traps. The more familiar people are, the more likely they are to procrastinate on discussing critical issues. Familiarity often sweeps underlying competition and contribution imbalances under the rug.

Trap	Manifestation	Consequence
Unquantifiable Contributions	It’s difficult to objectively measure who did the most work and who deserves first authorship.	The heavier contributor feels exploited, while the lighter contributor feels entitled.
Hidden Competition	Labmates are simultaneously competing for graduation timelines, grants, job placements, and the PI’s attention.	The partnership devolves into a covert rivalry.
Relationship Hostage-Taking	Feeling too awkward to formally discuss authorship order, funding distribution, or data rights.	Grievances bottle up and eventually explode.

We assume working with friends is easier, but the biggest hurdle is the awkwardness of speaking candidly. You feel awkward asking, “How are we handling authorship?” and they feel awkward admitting, “I actually don’t have much time for this.” Consequently, both sides rely on guesswork, which inevitably breeds misunderstandings.

The main threat in labmate collaborations isn’t a lack of closeness, but a total collapse of professional boundaries due to that closeness. A mature peer collaboration clarifies the professional stakes before leaning on personal rapport.

Documenting the division of labor, authorship expectations, data access, and deadlines isn’t an insult to the friendship, it’s an insurance policy for it.

2.2 Cross-Institutional Collaborations: Navigating Familiar Strangers

Partnering with other universities, external institutes, or international teams is vital for pushing research boundaries. These projects bring fresh data samples, novel methodologies, new contexts, and broader academic visibility—but they also introduce massive uncertainty. While they appear to be purely professional alignments, they rigorously test each collaborator’s interpersonal skills and cultural adaptability.

Dimension	Internal	Domestic/External	International
Communication Cost	Low	Medium	High
Baseline Trust	High	Must be built	Highly dependent on institutional frameworks
Credit Allocation	Handled via lab norms	Needs upfront negotiation	Demands binding legal agreements
Cultural Differences	Minor	Medium	Obvious
Decision-Making Pace	Rapid	Moderate	Slow/Bureaucratic

The challenge of external collaboration isn’t just geography; it’s navigating clashing institutional cultures:

Some groups thrive on rapid trial-and-error, while others require extensive bureaucratic approval.
Some fields rank authors strictly by contribution size, while others default to alphabetical order.
To some, “next week” means a hard seven-day deadline; to others, it’s merely a loose suggestion.
Some universities permit retroactive paperwork, while others strictly forbid handling data without prior legal clearance.

The most frequent misstep in cross-institutional work is assuming your home institution’s unspoken rules are universal laws. Therefore, before starting, you must ask not only “What are we researching?” but also “How will we execute it?”: How often will we meet? Where will the data live, and who has access? How will we attribute credit? Do our respective institutions have specific ethics board (IRB), IP, or contract requirements?

3. Why Are Contributions Always So Hard to Measure?

The most explosive topic in any partnership is rarely “who is smarter,” but rather “who did more.” The issue is that scientific contributions cannot be precisely clocked like factory shift work. Academic output blends highly visible labor with vital, yet invisible, investments.

3.1 The Visibility Gap in Academic Workload

Research efforts generally fall into two buckets:

Explicit Contributions: Easily documented tasks that naturally dominate authorship discussions: designing experiments, collecting data, writing code, building models, analyzing statistics, drafting the manuscript, and designing figures.
Implicit Contributions: Unquantifiable tasks that often dictate whether a project survives at all: early conceptual brainstorming, brokering access to resources or platforms, navigating data permissions, managing timelines, pivoting the research direction during critical meetings, and providing emotional support or conflict mediation within the team.

The crux of the problem is human nature: we instinctively overvalue our own invisible labor while discounting the invisible labor of others. You might believe you “provided the breakthrough idea,” while your partner remembers it as just a casual chat. You feel exhausted from “aligning institutional resources,” while they think you just made a group chat. You pride yourself on “steadily managing the project pipeline,” while they only respect the person who ultimately typed up the manuscript.

Evaluating scientific contributions is incredibly difficult because the work defies standardization. How much is a brilliant “shower thought” worth? Does navigating a bureaucratic nightmare to get data access merit authorship? Whose work was absolutely essential, and whose was just a nice bonus?

Because the scientific community lacks a universal standard for these nuances, collaborators must define their own. The easiest fix is to draft a “contribution checklist” at kickoff and update it dynamically as the workload shifts.

3.2 The Recency Bias in Recognizing Contributions

Many disputes arise from a severe recency bias: giving all the credit to whoever is sweating the most right before submission, while ignoring the heavy lifting done early on. A successful project isn’t built on a single “highlight reel” moment; it’s the sum of countless sustained efforts. Different types of labor are historically overvalued or undervalued depending on the project phase:

Phase	Highly Valued Work (Usually Rewarded)	Easily Undervalued Work (Often Ignored)
Ideation	Formulating the research question, defining the scope, building the theoretical framework, and designing the methodology.	Synthesizing literature, assessing field needs, and background research.
Execution	Solving technical bottlenecks, running experiments, and optimizing models.	Routine project management, chasing down teammates, and tedious data cleaning.
Publication	Drafting the text, finalizing data analysis, and generating polished figures.	Formatting to journal guidelines, running supplementary analyses, and crafting reviewer rebuttals.
Post-Publication	Giving conference talks, boosting academic visibility, and securing follow-up grants.	Archiving materials, maintaining external relationships, and communicating findings to the public/industry.

If credit goes solely to the final writer, the architects of the idea feel cheated. If all the glory goes to the “ideas person,” the researchers who actually executed the vision feel exploited.

A more equitable approach views the project as an integrated chain: ideation, methodology design, resource acquisition, execution, writing, and publication. Every link is vital. However, the relative value of each link fluctuates depending on the specific project, which is why upfront communication is mandatory.

3.3 Preventing Authorship and Contribution Disputes

Collaborations rarely implode because people are inherently greedy; they fail because of a lack of visibility. In the trenches of research, you are hyper-focused on your own sacrifices. You don’t see what I’m doing behind the scenes, and I don’t see the late nights you’re pulling. Consequently, everyone feels like a martyr.

Upfront Agreements: Lock down roles and tentative authorship on day one. At a minimum, clarify four things: who is doing what, what the specific deliverables are, the preliminary rules for author order, and the protocol for adjusting authorship if workloads change.
Early Dicussion: Never wait until submission week to discuss credit. Implement a monthly progress check-in and a quarterly contribution review. These aren’t interrogations; they are safety valves to identify if someone is overburdened, if someone is slacking, or if the initial division of labor requires recalibration.
Document Progress: Document important decisions, even with close friends. This can be done casually. For instance, drop a message in the group chat after a meeting: “Just to recap, A is handling the wet lab, B is running the stats, and C is drafting the intro. Does that sound right?” This isn’t a sign of distrust; it’s a safeguard against misremembering.

True fairness isn’t just about the final author order; it’s about maintaining transparent communication throughout the entire journey. When everyone’s hard work is visible and acknowledged, minor grievances evaporate. Conversely, if credit relies solely on foggy memories and hurt feelings, the final stages of a project will inevitably devolve into bitter arguments over “he-said, she-said.”

4. Navigating Collaborative Pitfalls in the Age of AI

Artificial Intelligence is radically altering the daily workflow of academic research. However, we must be absolutely clear: while AI drastically boosts efficiency, it cannot blur the lines of accountability. Tasks that previously demanded weeks or months—data cleaning, drafting code, literature reviews, and outlining—can now yield results in mere hours via AI. While this supercharges productivity and lowers the barrier for interdisciplinary work, it also introduces unprecedented hazards. The core dilemma is this: AI accelerates output, but it obscures who is actually responsible for that output.

Researchers are juggling multiple commitments. When contributing to a joint project, it’s highly tempting to lean on AI to save time. But when a partner uses AI to crunch numbers, write scripts, or draft text, a new form of “boundary testing” emerges: Does my partner actually understand the math behind this analysis? Has this code been rigorously debugged? Did the AI hallucinate these citations? Was sensitive proprietary data fed into a public language model? If ground rules aren’t established immediately, these AI shortcuts become massive liabilities.

4.1 Data Analysis Risks: AI is an Assistant, Not a Statistician

Teams must explicitly negotiate whether, and how, AI tools are permitted during the research process. While AI is excellent for brainstorming analytical frameworks, cleaning up code, suggesting visualizations, and drafting statistical interpretations, it introduces three severe risks:

Risk	Typical Manifestation	Potential Consequence
Misapplication of Methods	Blindly trusting AI to select statistical models without grasping the underlying data distribution.	Invalid statistics and scientifically flawed conclusions.
Over-interpreting Noise	Prompting AI to spin or over-explain statistically insignificant findings.	The narrative sounds convincing but is entirely unsupported by the actual data.
Data Breaches	Feeding raw, non-anonymized, or proprietary clinical/corporate data into public LLMs.	Severe violations of IRB protocols, NDAs, and institutional data security policies.

Total transparency is required when partners employ AI. This is particularly urgent when handling sensitive information like human clinical records, corporate trade secrets, or unpublished experimental results. The consortium must draw a hard line: exactly which datasets are safe to feed into cloud-based AI, and which must remain strictly confined to local servers or secure, compliant platforms.

Furthermore, rigorous AI-assisted workflows must be mandated. For instance, any data pipeline generated with AI assistance must preserve the raw data, the exact prompts used, the cleaning logs, and the analytical logic. Crucially, a human expert who deeply understands the methodology must manually verify the output.

4.2 Coding Risks: AI Writes Fast, But Fails Silently

AI coding assistants undeniably turbocharge productivity, instantly churning out scripts, functions, boilerplate comments, and unit tests. Yet, the pitfalls of AI-generated code are notorious: it often compiles and runs, but executes the wrong mathematical logic; it might work perfectly on a toy dataset but crash spectacularly in production; or the variable names look highly professional, but the core algorithmic steps are a black box.

Consequently, collaborators must be on high alert for the following red flags:

A partner pushes a block of code but cannot articulate how the core algorithm actually works;
There is no Git commit history, making it impossible to trace which version of the script produced the final figures;
Spaghetti code where data cleaning, model training, and plotting are all jumbled together, destroying reproducibility;
The AI “hallucinated” software packages, functions, or hyperparameters that don’t actually exist;
The AI quietly inserted aggressive data-filtering steps that artificially inflate the final statistical significance.

To combat this, teams must enforce strict “code accountability.” The rule is simple: whoever commits the code owns the responsibility for explaining it, testing it, and maintaining it. Whoever uses that code to generate a figure for the paper is on the hook for its reproducibility.

Best practices include: relying heavily on version control (like GitHub); mandating basic unit tests for critical mathematical functions; meticulously documenting the runtime environment and dependencies; and requiring a “code review” by a secondary co-author before submitting the manuscript.

4.3 Writing Risks: AI Can Polish Prose, But It Will Hallucinate Facts

AI is most commonly deployed in academic writing to smooth out clunky phrasing, structure outlines, draft abstracts, trim word counts, and brainstorm the discussion section. These use cases are generally benign. The critical danger lies in accepting AI-generated text as empirical truth. Severe risks include:

Hallucinating fake papers, fabricated DOIs, or phantom authors;
Presenting theoretical hypotheses from related literature as ironclad, proven facts;
“Spinning” the results so heavily that the discussion makes claims the raw data simply cannot support;
Unilaterally rewriting a co-author’s nuanced scientific argument via an AI prompt without their consent.

Failing to catch these AI hallucinations early practically leads a “Desk Reject”.

5. How to Break the Gridlock in Scientific Collaboration

An elite collaboration isn’t one devoid of conflict; it’s one equipped with mechanisms to resolve it. It doesn’t rely blindly on the “good character” of its members. Instead, it systematically rewards reliable contributors while rapidly exposing bad actors. We must upgrade mutual trust from a vague “gut feeling” into a codified, actionable system.

5.1 Draw Clear Professional Boundaries

Enforcing boundaries isn’t acting coldly; it’s installing guardrails to protect the partnership. You must broadcast your absolute limits, effectively shutting down any opportunity for partners to test your tolerance.

Category	Negotiable	Unacceptable
Time	Reasonable deadline extensions, proactive heads-ups, and organized rescheduling.	Chronic ghosting, missed meetings, and malicious stalling.
Tasks	Rebalancing the workload based on honest capacity limits.	Weaponized incompetence, passing the buck, and demanding ‘ghost’ authorship.
Resources	Sharing datasets and computing power precisely as negotiated.	Hijacking data for side projects or leaking proprietary materials to third parties.
Results	Reordering authors based on a transparent review of actual labor.	Intellectual theft, rogue journal submissions, or poaching the project.

Many academics dread establishing boundaries, fearing it makes them look “difficult” or stingy. In reality, boundaryless collaborations don’t breed intimacy—they breed exhaustion. Explicit boundaries define the “safe zone” for both parties, injecting much-needed predictability into the project.

5.2 Implement Checkpoints: Solve Small Problems Before They Snowball

The gravest threat to a project isn’t a mistake; it’s a mistake that is swept under the rug for months. Teams must build rigorous communication loops, scheduling mandatory “checkpoints” tailored to the project’s lifecycle:

Check Cycle	Check Content	Purpose
Monthly	Tracking progress, flagging roadblocks, and mapping next steps.	Catching minor execution errors before they compound.
Quarterly	Reviewing actual labor vs. expected labor, and recalibrating authorship expectations.	Ensuring both sides still feel the deal is fair.
Milestone Approaching	Manuscript drafting, journal submission, running reviewer-requested experiments, and finalizing IP.	Preemptively neutralizing bitter fights over final credit.

These checkpoints must be hardcoded into the calendar. Calling an “emergency meeting” only after tempers flare is a recipe for disaster. Routine check-ins drastically lower the emotional stakes, removing the defensive “Why are you attacking me?” reflex. Excellent project management isn’t about micromanaging your peers; it’s about forcing risks into the light early. A slightly awkward 15-minute chat in month three can prevent a scorched-earth screaming match in month twelve.

5.3 AI Pacts: Mandating Transparency to Prevent Algorithmic Black Boxes

Thriving in the modern era doesn’t mean banning ChatGPT; it means outlawing “shadow AI”—the secret use of tools with zero verification and zero accountability. Therefore, AI ground rules must be a standard part of the kickoff meeting for any co-authored paper. Demanding AI transparency prevents the formation of new “trust black boxes.” Teams should adopt three golden rules: AI is a co-pilot, not a co-author. The human assumes 100% liability; core text generated or heavily polished by LLMs demands strict manual auditing; and every single citation, data point, and scientific conclusion must be hand-verified.

Scenario	What you can do	What must be avoided
Data Analysis	Using AI to brainstorm statistical approaches, write boilerplate scripts, or suggest plotting libraries.	Dumping raw patient data into public web prompts or blindly trusting an AI’s P-value interpretation.
Code Design	Auto-generating syntax, adding docstrings, and building unit tests.	Copy-pasting “black box” algorithms directly into the production pipeline.
Manuscript Writing	Fixing grammar, outlining sections, or summarizing the intro.	Letting the AI invent fake DOI links or exaggerate the scientific implications.
Project Management	Condensing Zoom transcripts into action items.	Leaking embargoed, pre-patent intellectual property into unsecured cloud tools.

The most profound shift AI introduces isn’t speed; it’s the illusion of effort. Contributions manifest instantly, rendering actual labor invisible and highly untraceable. If a partner leverages AI to crank out a massive literature review or thousands of lines of code, the team faces a philosophical dilemma: How do we weigh this contribution? Who audits it? If the AI plagiarizes, whose career takes the hit?

Moving forward, kick-off meetings must explicitly cover AI boundaries alongside data rights and funding. The gold-standard collaborator of the 2020s isn’t a luddite who refuses to touch AI; it’s a professional who meticulously documents their prompts, rigorously audits their outputs, and unconditionally owns the final results.

5.4 Always Have an Exit Strategy: Leverage Requires Options

The most suffocating position in academia is being trapped in a toxic partnership because you literally cannot afford to leave. Your PI holds the data hostage, a peer is the only one who can run the critical machine, or your PhD defense is singularly tied to this one paper. Cultivating the ability to walk away isn’t about being a quitter; it’s about ensuring you negotiate from a position of strength, not desperation:

Diversify your portfolio: Never bet your entire publication record on a single joint venture.
Cultivate a wide network: Ensure no single gatekeeper controls your access to crucial lab equipment or datasets.
Control your IP: Maintain pristine, localized backups of your code, bench notes, and data so you can rebuild if cut off.

A stellar partnership should be a force multiplier, not a life raft. When both parties know they can survive independently, communication instantly becomes more honest and egalitarian. The bedrock of a healthy collaboration is that it remains a continuous, mutual choice between two highly capable, independent scientists.

5.5 Cultivating Long-Term Partnerships: Building a Track Record of Verifiable Trust

Researchers shouldn’t just optimize for a single Nature paper; they should be scouting for career-long allies. Evolving from a one-off gig into a “dream team” requires a massive accumulation of verifiable trust. A partner worth keeping around usually exhibits these traits:

They under-promise and over-deliver.
If they miss a deadline, they give a proactive heads-up instead of ghosting.
They loudly champion their teammates’ work and refuse to hog the spotlight.
When authorship disputes arise, they tackle them head-on in the daylight.
They honor data embargoes and IP agreements long after the paper hits the presses.

Decades-long alliances aren’t forged overnight; they are stress-tested through hundreds of minor interactions. For a scientist, the holy grail isn’t a partner who never makes a mistake; it’s a partner who, when the inevitable crisis hits, steps up, communicates transparently, fixes the bug, and takes full ownership of the fallout.

6. Conclusion

By embracing difficult conversations upfront, a collaboration actually has the legs to go the distance. Navigating academic partnerships is a masterclass in diplomacy; it forces us to synchronize with vastly different personalities, work speeds, and hidden agendas. It brutally reminds us that “good vibes” and unspoken rapport are laughably insufficient. Enthusiasm inevitably wanes, and “tacit understandings” are almost always misjudged. Only rigid professional boundaries, transparent accounting of labor, and relentless communication can bulletproof a project.

So, banish the fear of asking “awkward” logistical questions on day one. Demanding a clear division of labor isn’t being a micromanager; demanding authorship rules isn’t being ruthless; keeping a paper trail isn’t a sign of paranoia. On the contrary, these are the hallmarks of a consummate professional. Uncomfortable honesty protects the relationship. May every collaborative endeavor not only advance human knowledge but also elevate everyone involved, paving the way for a lifetime of shared discovery.

References

[1] Leon C, Lipuma J. Overcoming obstacles to interdisciplinary research: Empirical insights and strategies. Journal on Systemics, Cybernetics and Informatics. 2024.12,22(6):18-34.
[2] Newman J. Cultural barriers to interdisciplinary research collaboration: evidence from Australia. Humanities and Social Sciences Communications. 2025.11,12(1):1-3.
[3] Vantard M, Galland C, Knoop M. Interdisciplinary research: Motivations and challenges for researcher careers. Quantitative Science Studies. 2023.12,4(3):711-27.
[4] Moats D, Holtrop T, van Eck NJ, Varga J, Waltman L, Dechesne F. Making problems: interdisciplinary collaboration and AI ethics. Science as Culture. 2025.6,29:1-26.
[5] Gridach M, Nanavati J, Abidine KZ, Mendes L, Mack C. Agentic ai for scientific discovery: A survey of progress, challenges, and future directions. arXiv preprint arXiv:2503.08979. 2025.12.